ࡱ> X` objbj  !j6666 4j4j4j,kk%lnmnmnmnmќmPHJJJJJJ$h-n Y^n6fnmnm,Znm nmHHp  nmtl o94j{\PLH%h8׷0 \Cǣ˥nnj%(Dl)0H"(l0H666  Semantic and Machine Learning Methods for Mining Connections in the UMLS Research Proposal (for Committee Review) Chintan Patel 3rd Year Ph.D. Student Clinical Informatics Track Research Advisor: Dr. James Cimino Committee: Dr. Stephen Johnson Dr. Adam Wilcox Department of Biomedical Informatics Columbia University Specific Aims The Unified Medical Language System (UMLS) is an invaluable resource for the biomedical community. By integrating biomedical ontologies and terminologies that are used for annotation, organization, sharing and retrieval of biomedical data, the UMLS provides a fundamental resource that drives both basic biomedical discoveries and realization of the nationwide interoperable healthcare infrastructure. One of the intended uses of the UMLS Metathesaurus (Meta) is to support the translation of terms from a source terminology into terms in a target terminology. The concept orientation feature of UMLS provides a direct mechanism for translation - the source term corresponds to a concept that also has a corresponding term in the target terminology. More general types of "translation", as evident from the research literature on UMLS NOTEREF _Ref146371092 \h  \* MERGEFORMAT 2,,,, involve: (1) finding target terms with the closest meaning to the source terms (also known as "mapping"), (2) finding target terms that are related to the source terms (e.g. organisms to diseases caused by the organism) that can serve as proxies for the source terms for various functions (such as information retrieval or knowledge discovery), and (3) finding target terms that satisfy some semantic or structural constraint (e.g. information theoretic distance). The methods for finding such "translations" or connections between terms in Meta (other than the case of one-to-one synonymy) are not at all clear. In fact, the S in the UMLS the "system" - has not yet been fully characterized, let alone realized. Previous attempts NOTEREF _Ref146371176 \h  \* MERGEFORMAT 5, NOTEREF _Ref146371411 \f \h  \* MERGEFORMAT 6, to exploit such connections have depended on either manual selection of relevant connections, or problem-specific algorithms that use expert knowledge about the relative suitability of various inter-concept relationships. Where these attempts have succeeded, they have remained either labor intensive, difficult to maintain, or unscalable beyond their original domains. We believe that machine learning techniques offer automated, generalizable approaches that are appropriate for use with the UMLS, given the large set of potential connections and the need for a problem-independent approach. We are especially interested in solutions that exploit the respective advantages of logical and statistical reasoning. As we will describe in this proposal, there are fundamental challenges to using machine learning approaches with datasets such the Meta, including the complex relationships that exist in any potential training samples, and the relatively unexplored implications of semantic aspects in the data set (i.e. the conceptual and relational nature of the Semantic Network and Meta). In this proposal, we will describe prior work with methods such as statistical relational learning, probabilistic graphical models and social network analysis NOTEREF _Ref146378232 \h  \* MERGEFORMAT 43, that have provided some initial leads, and we propose ways to extend those methods for extracting connection pathways from the UMLS. We hypothesize that learning strategies that exploit the relational features, scale free properties and probabilistic dependencies of connections in the UMLS will identify meaningful inter-term relationships and that a combined approach will perform better across different problem domains when compared to any of the approaches in isolation. The specific aims of this project are to: Characterize the relational features in the UMLS and develop machine learning algorithms to identify appropriate term connections, based on the semantics, probabilistic dependencies and scale free property of the UMLS. Evaluate the proposed approaches with training datasets from a variety of problem domains. Implement algorithms into the UMLS Knowledge Source (UMLSKS) API toolkit for mining and visualizing the connections. There has not been, to our knowledge, any concerted effort to perform machine learning for mining connections over a large semantic resource such as the UMLS. While we have obtained some initial success with some of the proposed approaches, a generalized set of connection mining methods applicable to different problem domains remains to be explored. We believe that the UMLS provides a unique fertile ground to develop novel semantic relational mining methods and advance our understanding of mining large biomedical concept graphs. Furthermore, if our methods prove successful, they should be adaptable to a variety of applications that need to translate, or otherwise find connections among, terms across disparate terminologies, thus putting the "S" in "UMLS". Background and Significance In order to describe our proposed approach, we first provide some brief background information about the various machine learning techniques we will employ. We also describe the aspects of the UMLS that are relevant to the relational features that will serve as input to our methods, as well as prior relevant work that has studied extraction of connections from the UMLS. 1. Machine Learning History and Nomenclature The notion of "learning" mathematical models from given data to make future prediction can trace its origins to fundamental work by Karl Friedrich Gauss(1800s), Karl Pearson (1900s) and Ronald Fisher (1932), among others. More recently, researchers in the field of artificial intelligence (AI) seeking to develop intelligent learning systems (for diagnosing patients, cleaning floors, running cars, etc.), developed heuristic approaches. Soon, AI researchers began to discover underlying statistical bases for heuristics, spawning the sub-field of AI called machine learning. In late 1980s and early 1990s, with the growth of large scale medical and business databases, the notion of "mining" these repositories to discover "interesting" patterns led to the field of data mining. Additional methods have been developed to support data mining with real-world data, including data cleansing, data visualization and hypothesis testing. The study of these methods together with data mining is referred to as knowledge discovery in databases (KDD), with the main goal being the identification of useful knowledge from available data. In general, the machine learning or data mining approaches can be categorized into supervised and unsupervised methods. The cases where explicit training data (or labeled examples) are used to learn a model (classification or regression) for making future predictions fall under supervised learning approach whereas the cases where the algorithms learn a model from unlabeled data (without any notion of output) belong to unsupervised approach. Another recent approach, relevant to this proposal, includes semi-supervised learning where the classifier is constructed using both labeled and unlabeled examples. 2. Relational Data Mining In early 1990s it became evident that the "single table assumption" used in mining databases didnt apply to real world datasets where the data were distributed across multiple normalized tables (e.g. patient demographics and lab data). Naively applying standard machine learning methods over joins or aggregated dataset led to loss of information and incorrect results. The field of relational data mining10 approached the problem by using methods from Inductive Logic Programming (ILP) to learn over such datasets. The ILP methods model the data, schema and background knowledge as first order logic predicates; generalized rules are inductively learned from these facts, which can then be applied on future instances. The methods from relational data mining have been successfully applied in molecular structure activity prediction, constructing biological pathways, etc. 3. Statistical Relational Learning More recently, a variety of large scale network datasets have become available electronically, for example, social networks, the World Wide Web, terrorists network, metabolic and other biological networks. The focus has thence shifted to relational learning approaches that are more statistically oriented (rather than logic-based) and are broadly grouped under the domain of statistical relational learning(SRL).The core underlying assumption in SRL is that the training examples are not independent and identically distributed (i.i.d) but related in complex network fashion. There has been convergence of methods from probabilistic learning NOTEREF _Ref149220859 \h  \* MERGEFORMAT 27, web mining, social network analysis NOTEREF _Ref149220900 \h  \* MERGEFORMAT 44 to mine the aforementioned datasets for predicting protein-protein interactions, ranking web pages, finding links between friends or terrorist, etc. These applications can be broadly grouped in to learning tasks such as link classification, link prediction and group detection, which we describe in detail in the next section. The presence of links across data instances requires careful consideration of statistical issues while mining connections. The methods in this domain are still evolving and the networks studied are relatively poor in the variety of semantics of nodes or edges when compared to Meta. An interesting analysis shows how the presence of semantic types for nodes and edges can be exploited for relationship detection applications. Another effort relevant to our work is the semantic association ranking approach; the authors use an association ranking score that is based on semantics of relationships, concept depth, etc, for mining connections across semantic entities. The approach has limited applicability since it does not use any learning paradigm. 4. General Classes of Connection Mining Problems We now describe the general classes of problems involved in mining connections and later discuss how these problems instantiate to specific connection mining problems in Meta Link Classification: A set of connection paths that are judged as "correct" (the training set) is used to train the system to classify\label additional connections (the test set) in supervised manner. Once classified, links can be used to classify the nodes in the links, for example, determining the category of a web page based on the outgoing hyperlinks to a set of labeled web pages. Connection Clustering/Community Identification: In this case we are interested in identifying hidden clusters of nodes based on the connections they are involved in (hence connections are used as features). This set of problem appears in uncovering virtual communities within social networks. Link Prediction: The goal is to predict whether a connection exists between two nodes and determine the strength of the connection based on some score. This has been traditionally used for studying terrorism networks and protein interactions. Collective Inference: In network datasets, defining separate training and testing sets is difficult, since all the entities are typically connected in the network. The problem of collective inference deals with situations wherein the values/labels of a few nodes in the network are known and the goal is to determine labels of other all other nodes in the network. This problem class is also relevant in semi-supervised settings, where only positive training data are available. One prominent use-case of collective inference is studying regulatory networks of cell. Several of the connection mining problems in Meta become special instances for these general problem classes. The terminology translation can be modeled as link prediction or link classification problem as described in the preliminary work section. The problem of clustering phenotype concepts based on their links to genotype concepts2 is an instance of connection clustering application. The link prediction can be used for finding connections suitable for information retrieval tasks such as linking patient data to information resources and also for knowledge discovery applications. 5. Probabilistic Graphical Models: Bayesian Networks One method for modeling and inferring connections in complex structures such as the UMLS is probabilistic graphical modeling, which integrates probability theory and graph theory. Several other important biomedical problems,,,, can be easily modeled using this framework. Bayesian networks or belief networks are one of the most widely studied instantiation of probabilistic graphical models. The core idea behind Bayesian networks is simple: given a directed acyclic graph, the nodes represent random variables and the edges (or arcs) represent the probabilistic dependency among the variables. The most intriguing feature of Bayesian networks is that the joint probability distribution of all random variables can be compactly represented as shown below p(X1,X2,X3Xn) =  EMBED Equation.3  where the joint probability of random variables X1 to Xn is product of conditional probability of Xi given the parents of Xi . This simplification models the conditional independencies present in the model (e.g. consider a simple Bayesian network: high temperature ( fever ( flu, the presence of fever, makes flu independent of high temperature). Some current research challenges include selecting the best Bayesian network structure (from all possible permutations) that maximizes the likelihood of observed data and estimating the conditional probability distribution for variables when the training data is not complete21. One of the promising applications of learning Bayesian network structure over large graph-based datasets uses association rule mining (counting frequent items that occur often together) to model co-authorships networks. Once the network is constructed, the next step involves inference or answering queries wherein we are given values for a set of variables (evidence) and the goal is to find probabilities for other variables in the network. To handle tractability issues, the inference algorithms tend to trade off accuracy for efficiency. Bayesian networks can be used to perform connection mining, as demonstrated with the development of Probabilistic Relational Models (PRM). The essential idea behind a PRM is that a random variable can considered as an object (in the object-oriented sense) that can have attributes which are themselves random variables with probabilistic dependencies across other objects. PRMs have been successful in representing Bayesian networks over relational databases by directly modeling database tables as objects, the columns as attributes and the key constraints as attribute dependencies. However, it is not clear how PRMs can be used over network or graph datasets where there is no distinct notion of schema and instances. 6. Support Vector Machines Support Vector Machines (SVMs) embody a supervised learning approach used for classification and regression problems. The groundwork leading to SVMs can be traced to the statistical learning theory developed in 1970s due to Vapnik and Chervonenkis. Recently, SVMs have been receiving increased attention which can be attributed to two of their interesting properties. Firstly, SVMs perform structured risk minimization, which implies that of all possible solutions (i.e. the planes separating positive and negative samples), SVMs try to find the best solution whereas traditional algorithms such as neural networks stop when they find the first solution. Secondly, SVMs use kernels to map the data input into a higher dimensional feature space where the dataset becomes linearly separable. More concretely, the kernel, k, is a function that takes as input data vectors and produces a scalar value by performing a dot product,  EMBED Equation.3  where, x and y are input vectors and ( is the mapping function that maps the input values to an expanded feature space. The most important property of kernels (the kernel trick) is that they calculate the dot product without expanding the mapping functions ((). In the overall classification or regression function, f(x) =  EMBED Equation.3 , the parameters, ( (support vectors) and b (bias) are learnt from the training data. One of the applications of kernel-based methods for connection mining paradigm uses concept graphs as the input vectors. Recently, SVMs have been shown to be closely related to the classical regularization theory and spectral graph theory. Promising applications of these theories for semi-supervised learning and ranking over graph-based data are being actively researched. Learning over semantically typed graphs such as Meta remains relatively unexplored in the machine learning community. 7. Social Network Analysis and Scale Free Networks The field of social network analysis (SNA) provides methods to study the social relations among a set of actors (people, groups or organizations). The relations are primary object of analysis, the attributes of actors are not generally considered. SNA considers various properties of networks, such as betweenness, centrality, clustering coefficient that are directly applicable for the proposed connection mining problem over Meta (see section D5). An important discovery from SNA (due to Barabasi et al) that has implications beyond social networks is that most real world networks are scale-free - there are few nodes with a high degree of connectivity (hubs) with most other nodes have a low degree of connectivity. Some examples of scale free networks are the Internet, the World Wide Web, the Hollywood actor network (Kevin Bacon being the most famous example), metabolic networks etc. One of the ways of detecting scale freeness is to plot a log-log plot of following power law equation: p(k) ( k-( which says that the probability, p(k), of a node connected to k other nodes is proportional to k-( where ( is any constant. More recently a precise metric for measuring scale-freeness was proposed by Li et al. Another related term is that of small-world networks, (closely related to the notion of six-degrees of separation) which says that the average path distance between any two nodes in the network is very small. The small-world property of a network is generally implied by the presence of scale free structure. Analyzing the network structure to identify modular structures responsible for specific function or property has recently led to range of novel discoveries,,. 8. The Unified Medical Language System The Metathesaurus (Meta) of the Unified Medical Language System1 (UMLS) is a collection of terminologies in which synonymous terms (from the same different terminologies) are mapped to single concepts, each of which has a Concept Unique Identifier (CUI). The concepts are related to each other through relationships largely obtained from the source terminologies. The semantic types themselves are organized into a hierarchical structure that includes relations among the types. The 2006AC version of Meta integrates 138 biomedical terminologies and there are currently total 1,352,403 unique concepts in the Meta. The SN provides top-level concepts (e.g. Clinical Drug, Laboratory Test) to categorize the concepts in Meta. The relationships present in the source terminologies are asserted between the concepts (and not between the terms) in Meta, hence we obtain relationships across different terminologies as shown in the schematic in Figure 1. The UMLS is updated and disseminated three to four times every year as a set of files. We will be using following files: MRSTY (for concepts and corresponding Semantic Types), the MRREL (for the relationships and corresponding source terminology), and the MRCONSO (concepts and their source terminologies). Figure 1: A schematic showing how the relationships from a single source terminology span across different terminologies. 9. Definitions of Meta Features We define source concepts and target concepts as being a Meta Concept to which terms from a source or target terminology (respectively) are mapped. We define a connection as a set of transitive relationships between a source concept and a target concept connected by zero or more intermediate terminologies through links in MRREL. Consider for example, given the terminologies in Figure 1, Verotoxigenic E. Coli gastrointestinal Infection (ICD-9-CM) is connected to Verotoxin 1 (MeSH) through E. Coli 0157 (SNOMED-CT). 0-Step Connection A loop relationship between terms from different terminologies that are "synonymous" that is, the source concept and the target concept are identical. 1-Step Connection A direct relationship(s) asserted in the UMLS file MRREL between a source concept and a different, unique target concept. 2-Step Connection A set of two transitive relationships, wherein the source concept is related to a second concept (with terms from some other terminologies), which in turn is related to the target concept. Similarly, one can envision n-Step connections that consist of n transitive relationships across n+1 concepts. We define a connection set as all possible connections between a set of source concepts and a set of candidate target concepts. 10. The UMLS and Connection Mining When addressing the need for inter-concept connections, UMLS researchers have traditionally focused on deterministic algorithms that use some domain or expert knowledge for discovering cross-terminology paths. Cimino et al.5 first proposed an algorithmic approach to convert ICD-9 to MeSH terms using the UMLS. Although, the authors found links between 95% of concepts between ICD-9 and MeSH, they rightly concluded that the methods were not clear to select the "best" translations or to generate an expression consisting of multiple target terms. The current proposal will address the need for discriminatory method by using machine learning approaches to rank and select from a set of possible connections the connection most appropriate for a given problem and pair of terminologies. Bodenreider and colleagues6 used a deterministic graph traversal algorithm to exploit the "semantic locality" in UMLS to map any concept in the UMLS to MeSH. Zang and colleagues described an approach for using associative relationship patterns to map different ontologies. These associative patterns were "mined" using simple frequency counts based on occurrences, but no methodical approach was used to measure the support or confidence for the patterns. More recently, Fung et al.9 showed another deterministic approach that uses the synonymy, ancestor expansion and descendant expansion methods over SNOMED-CT to ICD-9-CM mappings as gold-standard. To our knowledge, there have been no publications to date (other than our own, discussed in the Preliminary Studies section, below) describing the use of statistical relational approaches for mining cross-terminology connections from the UMLS. However, some UMLS applications have implicitly used some form of connection mining. In their recent work on phenotype analysis, Butte et al2. used the relationships in Meta to create a genome-phenome network. A terminological approach to discovery of gene and disease relationships was shown in Genestrace application NOTEREF _Ref146378610 \h  \* MERGEFORMAT 7 by traversing the relationship paths in the UMLS. Various projects, dealing with ontology integration and maintenance often can be solved using connecting mining approaches. Existing methods generally rely on simple lexical matches, structural matches, or deterministic path traversal algorithms largely driven by domain specific heuristics. 11. Notes on terminology used in this Proposal Given that the proposed methods (originating from different disciplines) are still evolving in the research community, the set of notations or terms used to describe the methods or informational entities might seem confusing to the reader. We provide some explanation of terms that will be used interchangeably based on the context. The terms mining, learning and machine learning imply use of some supervised, unsupervised or semi-supervised approach to finding patterns from the data. The terms relational mining, connection mining, statistical relational learning will be used to mean the class of machine learning approaches where the instances dont have the i.i.d (independent and identically distributed) assumption. The terms connections, translations, links, paths and pathways generally mean a transitive chain of the form : concept1, relationship1, concept2, relationship2, concept3 The terms graph, network will imply Meta; correspondingly, concept is equivalent to node and relationship to edge. 12. Significance With nearly 1,400,000 concepts and 22,000,000 relationships from about 140 terminologies, the Unified Medical Language System has grown into an enormous biomedical computational resource. The UMLS is now used in a variety of domains ranging from basic biomedical sciences to building national healthcare infrastructure. Existing efforts using the UMLS are only scratching the surface - the situation aptly illustrated by the Indian proverb about reconstructing an elephant by touching only a part, the leg makes it a tree trunk, the tail makes it a snake. We believe that using a machine learning approach, more specifically, relational connection mining is a promising direction to realize full potential of the UMLS for supporting inter-terminology translation. The UMLS is also uniquely positioned to serve as a testbed for understanding, evaluating and advancing the development of relational connection mining methods that could be applicable to any large network. Nearly two decades of largely manual efforts have been poured into the construction and maintenance of the UMLS. The rich semantics of concepts and relationships in the UMLS demand special consideration when developing statistical learning approaches over them. The important challenge lies in developing methods that can coherently integrate the statistical and semantic approaches to mine the connections. Furthermore, the network structure of UMLS also requires careful investigation; the concept hubs that arise due to the scale free property of the Meta network have important implications in generating the connection paths. Another aspect is concerned with modeling and inferencing over the probabilistic dependencies over the large graph such as the Meta. Given the range of connection mining problems, we propose using a combined strategy that takes into consideration all aforementioned important characteristics for mining connections in the UMLS. With the widespread pervasiveness of applications using the UMLS or its constituent terminologies, we believe that notion of using connection training sets is not a very far fetched assumption. One of the goals of the proposed approach is to allow end-users to input training connections (supervised or semi-supervised) or their datasets annotated with terminology concepts and then query for concept connections in the UMLS. Alternatively, we will provide users with a library of connections datasets and methods with flexible semantic constraints to query for connections. To bring the proposed approaches closer to the UMLS users, we propose to extend the UMLS KS (Knowledge Source) API to perform connection mining. Preliminary Studies 1. Recovering SNOMED-CT to ICD-9-CM Connections We used the dataset used in Fung et al9 to evaluate the possibility of using machine learning approaches for terminology translation. We removed the SNOMED-CT to ICD9CM translations from the UMLS 2005AA (where RELA= mapping_from or mapping_to) originally contributed from the SNOMED International. After removal of 72,976 such explicit edges, we found that 11,778 terms (16.1%) had the same CUIs, 35,452 (48.6%) translation paths were at 1-step distance, 24,199 (33.1%) at 2-step distance and rest (2.1%) had more than 2-step distance. For that work, we used a relational-features approach wherein we used semantic path features. We trained a Nave Bayes classifier using the "UMLS Sampling gold-standard - randomly sample the UMLS and consider 1-step connections as positive and strictly 2-step connections (that is, cases where a 2-step, but not a 1-step, connection existed) as negative. The algorithm was then evaluated over the 1-step and 2-step SNOMED-CT to ICD9CM connections, which resulted in positive classification of 99.82% of 1-step translation paths and 85.68% of translation paths with 2-step connections (see Appendix-1 for details on the experiment). We performed cross-validation for 2-step translation paths by mixing them with random 2-step strict paths (negative) and we were still able to get an AUC (area under the ROC curve) score of 0.94. The limitation of this approach is the assumption that we are given both the source and target concept. We dropped the assumption and tried to find all possible 2-step paths to ICD9CM from given source SNOMED-CT concepts but this resulted in a billions of potential translational paths, which we later attributed to the scale-free property of the UMLS network. We believe that the proposed approach of using scale free abstractions and semantic traversal will potentially solve the problem of exponential connection explosion. 2. EHR-based Information Retrieval and Connection Mining In our previous work, we showed how the UMLS terminologies can be used to support contextual information retrieval in electronic health records for Infobuttons. To show the applicability of connection mining for the problem of linking electronic health records and information resources, we used the OHSUMED information retrieval gold standard. The gold standard essentially contains 106 physician information needs and corresponding expert selected MEDLINE articles. We parsed the information needs using the UMLS tool MMTx to get SNOMED-CT concepts and extracted MeSH headings from the corresponding MEDLINE articles. We constructed gold-standard connections from 385 SNOMED-CT concepts to 2252 MeSH concepts. We then found 2-step test connections for the 385 source SNOMED-CT concepts to all possible MeSH concepts which where then classified using a relational features approach (semantic path). We were able to recover 94% of gold-standard MeSH concepts after classification. An important lesson learnt from the experiment was that several biomedical informatics problems/datasets can be modeled and solved using the connection mining framework. To show the general applicability of proposed methods, we will use a number of gold-standard training connections and evaluate them with different classes of connection mining problems. 3. The Scale Free Property of the UMLS We investigated the UMLS network structure using methods from social network analysis to understand whether Meta exhibits scale-free properties. We evaluated the source terminologies separately and Meta in its entirety. The results showed that although some terminologies didnt follow the power-law consistently, the overall Meta structure does exhibit scale-free characteristics as depicted in the Figure 2 with few concept hubs grabbing most of the relationships and large number of concepts connected by only a small number of relationships.  SHAPE \* MERGEFORMAT  Research Design and Methods In this research, we will develop and investigate approaches for mining connections in the UMLS. The overall workflow and major components of the grant proposal are sketched in Figure 2. Given the complexity of mining relational connections over the UMLS, we hypothesize that an approach that combines the following strategies will perform better than any single approach in isolation across different problem domains:- a). Relational Features: Develop or use features that capture or summarize the properties of the connection network and use them with standard machine learning algorithms. b). Scale Free Abstraction: Exploit the power-law property of the UMLS to abstract the network into a Hub Graph and mine the abstract graph structure to generate connections. c). Probabilistic Graphical Inference: Learn a Bayesian network structure from training connections and use semantic traversal to prune the number of concepts in the network to perform inference.  1. Training Connection Datasets An important aspect of the proposed research is to use or develop a wide range of training datasets and also to ensure the applicability of algorithms over a variety of connection mining problems that arise in the biomedical domain. Given, the widespread use of the UMLS based applications in the biomedical informatics domain, we believe that we can generate training connections from a several existing datasets for example, OHSUMED for modeling information retrieval type connections NOTEREF _Ref149228353 \h  \* MERGEFORMAT 52, SNOMED2ICD-10 mappings for learning translation connections, MeSH co-occurrence for literature based connections and so on. 2. Relational Features One of the approaches in the field of connection mining is to develop relational features that capture the relational characteristics of the dataset. The essential idea is that once we have calculated the features or "motifs" characterizing the connection paths, the i.i.d assumption resurfaces and standard machine learning methods can be then applied. The UMLS has several properties that allow development of such relational features, for example, number of relationships, concept degree, semantics of concept and relationships, etc. 3. Scale Free Abstraction Relational Learning Given the presence of large hub concepts having several relationships to other concepts make the UMLS a "small-world". The implications of this observation are critical in developing relational mining algorithms over the UMLS. We propose a novel Scale Free Relational Learning (SaFREL) approach that gives special consideration to the hubs in performing the learning and connection dataset generation. We will model a UMLS hub graph (HuG) and use this hub abstraction to develop novel modular mining algorithms. 4. Probabilistic Graph based Learning We will use the Bayesian network approach to model probabilistic dependences across the concepts in the UMLS. The eventual goal is to answer queries on the dependency graph of the form, Given the Concept A from UMLS, what is the probability of finding a meaningful connection to Concept B? We propose the notion of performing Semantic Traversal in the UMLS graph to scoup the portion of UMLS relevant to the training data and hence avoid scalability issues for inferencing. The parameters for Semantic Traversal will be calculated using the maximum likelihood approach. 5. Connection Ranking Given the fact that all proposed algorithms will result in some form of classification error score or connection prediction score for the test connections, we can use these scores to rank the test connections and generate the top k list for each method. We propose to develop a Connection Intersection Function(CIF), that performs an intersection of the top k lists using a relational approach. 6. Evaluation The evaluation of the proposed approaches is relatively straightforward, largely driven by the range of training connections/gold-standard datasets. We will use cross-validation approach to calculate a reliable metric such as the area under the ROC curve (AUC). The hypothesis will be considered proven if the CIF consistently results in higher overlap with gold-standard than individual methods across all training datasets. 7. Dissemination We propose to extend the UMLS Knowledge Source API NOTEREF _Ref146378454 \h  \* MERGEFORMAT 13 to support functions for mining connections from the UMLS. We believe that this is the best approach to disseminate the results of our investigation to the community at large. We will use a batch upload process and job request queues to allow transferring training connections and executing algorithms. 8. Milestones and Timeline a) Year 1 Milestones Develop training datasets and implement UMLSGraphGen (1st quarter) Implement and evaluate the relational feature-based algorithms (2nd quarter) Develop and evaluate Regularization and Kernels (2nd quarter - 3rd quarter) Implement UMLS HuG Generation (3rd quarter) Evaluate UMLS HuG traversal heuristics (3rd - 4th quarter) Summarization and publications of findings (1st 4th quarter) b) Year 2 Milestones Develop and evaluate Modular Mining (1st quarter) Implement Semantic Traversal for Meta (2nd quarter) Develop and evaluate probabilistic inference approach (2nd - 3rd quarter) Evaluate all approaches with CIF (3rd quarter) Implement ConnectionViz and UMLSKS API interfaces (3rd - 4th quarter) Summarization and publications of findings (1st 4th quarter) E. References     . Lindberg DA, Humphreys BL, McCray AT. The Unified Medical Language System. Methods Inf Med. 1993 Aug;32(4):281-91. . Butte AJ, Kohane IS. Creation and implications of a phenome-genome network. Nat Biotechnol. 2006 Jan;24(1):55-62. . National Health Information Infrastructure [home page on the Internet]. [cited 2006 Sept 12]. Available from: http://www.aspe.hhs.gov/sp/nhii/index.html. . Nelson SJ, Tuttle MS, Cole WG, Sherertz DD, Sperzel WD, Erlbaum MS, Fuller LL, Olson NE. From meaning to term: semantic locality in the UMLS Metathesaurus. Proc Annu Symp Comput Appl Med Care. 1991;:209-13 . Cimino JJ, Johnson S, Peng P, Aguirre A. From ICD9-CM to MeSH using the UMLS: a how-to guide. Proc.Annu.Symp.Comput.Appl.Med.Care. 1993:730--734. . Bodenreider O, Nelson S, Hole W, Chang H. Beyond synonymy: exploiting the UMLS semantics in mapping vocabularies. Proc.AMIA.Symp. 1998:815--819. . Cantor M, Sarkar I, Bodenreider O, Lussier Y. Genestrace: phenomic knowledge discovery via structured terminology. Pac.Symp.Biocomput. 2005:103--114. . Caviedes JE, Cimino JJ. Towards the development of a conceptual distance metric for the UMLS. J Biomed Inform 2004 Apr;37(2):77--85. . Lord PW, Stevens RD, Brass A, Goble CA. Investigating semantic similarity measures across the Gene Ontology: the relationship between sequence and annotation. Bioinformatics 2003;19(10):1275--1283. . Fung KW, Bodenreider O. Utilizing the UMLS for semantic mapping between terminologies. AMIA Annu Symp Proc 2005:266--270. . Dzeroski S. Relational Data Mining, Secaucus, NJ, USA: Springer-Verlag New York, Inc.; 2001. . Pearl J. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, San Francisco, CA, USA: Morgan Kaufmann Publishers Inc., 1988. . UMLS Knowledge Sources (UMLSKS) [home page on the Internet]. [cited 2006 Sept 8]. Available from: http://umlsks.nlm.nih.gov. . Carbonell JG, Michalski RS, Mitchell TM. Machine Learning: A Historical and Methodological Analysis. AI Magazine 1983;4(3):69-79. . Frawley W, Piatetsky-Shapiro G, Matheus C. Knowledge Discovery in Databases: An Overview. AI Magazine Fall 1992, pp. 213-228. . Chapelle O, Schlkopf B, Zien A. Semi-Supervised Learning, MIT Press, Cambridge, MA;2006 . Muggleton S. Inductive logic programming. New Gen. Comput 1991; 8(4):295--318. . Page D, Craven M. Biological applications of multi-relational data mining, SIGKDD Explor. Newsl 2003;5(1):69--79. . Hendrik B, David J, Stefan K. Introduction to the special issue on multi-relational data mining and statistical relational learning, Machine Learning 2006;62(1):3--5. . Chakrabarti S. Mining the Web: Discovering Knowledge from Hypertext Data, Morgan-Kauffman;2002 . Neville J, Simsek O, Jensen D. Autocorrelation and Relational Learning: Challenges and Opportunities. Proceedings of the Workshop on Statistical Relational Learning, 21st International Conference on Machine Learning 2004. . Barthelemy M, Chow E, Eliassi-Rad T.Knowledge Representation Issues in Semantic Graphs for Relationship Detection. In Proc of AAAI Spring Symposium 2005:91--98. . Meza BA, Wiener CH, Arpinar IB, Ramakrishnan SA. Ranking Complex Relationships on the Semantic Web, IEEE Internet Computing 2005; 9(3):37--44. . Getoor L. Link mining: a new data mining challenge. SIGKDD Explorations, 2003; 5(1):84--89. . Swanson DR. Fish oil, Raynaud's Syndrome, and undiscovered public knowledge. Perspect. Biol. Med 1986; 30: 7-18. . Husmeier, D. , Dybowski, R. & Roberts, S (eds.). Probabilistic Modeling in Bioinformatics and Medical Informatics (Springer, London 2005). . Cooper GF, Herskovits, E.: A Bayesian method for the induction of probabilistic networks from data. Machine Learning 1992, 9:309-347. . Friedman N. Inferring cellular networks using probabilistic graphical models. Science 2004;303 (5659):799-805. . Cooper G, Dash D, Levander J. Wong WK, Hogan W, Wagner M. Bayesian Biosurveillance. In M. M. Wagner, A. W. Moore and R. A. Aryel (Eds.), Handbook of Biosurveillance. New York City, NY: Academic Press;2006. . Chapman WW, Haug PJ. Bayesian modeling for linking causally related observations in chest X-ray reports. Proc AMIA Symp. 1998;:587-91. . Beaumont MA, Rannala B. The Bayesian Revolution In Genetics. Nat Rev Genet 2004;5(4):251--261. . Ogunyemi O, Clarke JR, and Webber B. Using Bayesian networks for diagnostic reasoning in penetrating injury assessment. In Proceedings of the Thirteenth Annual IEEE Symposium on Computer-Based Medical Systems; Houston, Texas;2000:115-120. . Heckerman D. A Tutorial on Learning Bayesian Networks. Microsoft Research Technical Report. MSR-TR-95-06. 1995. . Goldenberg A, Moore A. Tractable Learning of Large Bayes Net Structures from Sparse Data. In Proc of ICML 2004. . Agrawal R, Srikant R. Fast Algorithms for Mining Association Rules. In Jorge B. Bocca; Matthias Jarke & Carlo Zaniolo, ed.,'Proc. 20th Int. Conf. Very Large Data Bases, VLDB', 1994:487--499. . Needham CJ, Bradford JR, Bulpitt AJ, Westhead DR. Inference in Bayesian networks. Nat. Biotech. 2006;24(1):51--53. . Frey BJ. A comparison of algorithms for inference and learning in probabilistic graphical models. IEEE Trans Pattern Anal Mach Intell 2005;(9):1392--1416. . Friedman N, Getoor L, Koller D. Learning Probabilistic Relational Models. In Proc of IJCAI 1999:1300--1309. . Vapnik VN. The Nature of Statistical Learning Theory. Spinger-Verlag, Berlin, Germany; 1995. . Vapnik VN, Chervonenkis AY. On the uniform convergence of relative frequencies of events to their probabilities. Theory of Probability and Its Applications. 1971;16:264--280. . Burges, CJ. A tutorial on Support Vector Machines for pattern recognition. Knowledge Discovery and Data Mining 1998;2(2):121--167. . Cumby C, Roth D. Kernel methods for relational learning. In Proc. of ICML 2003. Washington, US. . Evgeniou T, Poggio T, Pontil M, Verri A. Regularization and statistical learning theory for data analysis. Comput. Stat. Data Anal. 2002;38(4):421--432. . Wasserman S, Faust K. Social Networks Analysis: Methods and Applications. United Kingdom: Cambridge University Press;1994 . Barabasi A, Albert R. Emergence of scaling in random networks. Science 1999;286(5439):509--512. . Li L, Alderson D, Tanaka R, Doyle JC, Willinger W. Towards a Theory of Scale-Free Graphs: Definition, Properties, and Implications. Internet Mathematics 2005. . Ravasz E, Somera AL, Mongru DA, Oltvai ZN, Barabasi AL. Hierarchical organization of modularity in metabolic networks. Science 2002 Aug 30;297(5586):1551-5. . Ma HW, Zhao XM, Yuan YJ, Zeng AP. Decomposition of metabolic network into functional modules based on the global connectivity structure of reaction graph. Bioinformatics 2004 Aug 12;20(12):1870-6. . Eom YH, Lee S, Jeong H. Exploring local structural organization of metabolic networks using subgraph patterns. J Theor Biol. 2006 Aug 21;241(4):823-9. . Zhang S, Bodenreider O. Comparing associative relationships among equivalent concepts across ontologies. Medinfo. 2004;11(Pt 1):459-66 . SNOMED-CT Mappings, http://www.snomed.org/clinical/SNOMEDMapping.html, Accessed 08 Sept. 2006. . Patel CO, Cimino JJ. Mining Cross-Terminology Links in the UMLS, AMIA Annu Symp Proc. 2006 (in press) . Lee Y, Supekar K, Geller J. Ontology integration: experience with medical terminologies. Comput Biol Med. 2006 Jul-Aug;36(7-8):893-919 . Noy NF. Semantic Integration: A Survey Of Ontology-Based Approaches. SIGMOD Record 2004;33(4):65-70. . Ling CX, Huang J, Zhang H. AUC: a Statistically Consistent and more Discriminating Measure than Accuracy. In Proceedings of the Eighteenth International Joint Conference on Artificial Intelligence 2003, Acapulco, Mexico, August 9-15:519-526. . Cimino JJ, Elhanan G, Zeng Q. Supporting infobuttons with terminological knowledge. Proc AMIA Annu Fall Symp. 1997;:528-32. . Hersh WR, Buckley C, Leone TJ, Hickam DH. OHSUMED: An Interactive Retrieval Evaluation and New Large Test Collection for Research. Proc of the 17th Annual International ACM-SIGIR Conference. Dublin, Ireland, 3-6 July 1994:192--201. . MetaMap Transfer (MMTx), Version 2.4.B, http://mmtx.nlm.nih.gov/, Accessed 12 Sept 2006. . Ling CX, Huang J, Zhang H. AUC: a Statistically Consistent and more Discriminating Measure than Accuracy. In Proceedings of the Eighteenth International Joint Conference on Artificial Intelligence 2003, Acapulco, Mexico, August 9-15:519-526. ICD SNOMED-CT MeSH Connection through relationships asserted in ICD Connection through relationships asserted in SNOMED-CT UMLS Concept Terms from Source Terminologies ICD: International Classification of Disease SNOMED-CT: Systematized Nomenclature of Medicine-Clinical Terms MeSH: Medical Subject Headings  Concept hubs  Figure 2: Analyzing scale freeness of the UMLS a). The power law plot (loglog) of number of concepts with k relationships and probability, p(k). b). A slice of 2,000 records from the MRREL file of the UMLS distribution depicting hub concepts and the gray scale indicating the number of source terminologies contributing the terms for the concept. a b Figure 2: The overview of major components involved in implementing and evaluating the proposed hypothesis. Relational Kernels (relSVM) Evaluation Mining Results Connection Ranking Connection Mining UMLSKS API ConnectionViz UMLSGraphGen UMLS Metathesaurus Probabilistic Graph (relBN) Scale Free Abstraction (SaFREL) Relational Features Training Datasets SNOMED-ICD Dxplain:Lab-Disorder PubMed abstracts ...  5TVghijn# $ % & + c d e s t Ľ}ud`Xh8ICJaJh hhhhCJOJQJ^JaJh8Ihh5h8IhhCJH*aJh8IhhCJaJ h8Ih93h93h(dhhCJaJh(dCJaJh(dh(dCJaJ h8Ihhh93CJ aJ h8IhoCJ aJ #hhho5CJ(OJQJ^JaJ(hh5CJ(OJQJ^JaJ(hhho  05TUVi$a$gdon  $ % & zUkd$$Ifl0+& t.'644 la & F m$If^gd8I $Ifgdhgdh$a$gdo& ' ( ) * + P d e s t T/gdh$a$gd8I $ & Fa$gd8I $ ` a$gd8I & F ^`gdh$a$gdh$a$gdot   !"2345dzǓrh8Ihh0JCJaJ&jh8IhhCJH*UaJ&jh8IhhCJH*UaJh8Ihh6CJaJ&jh8IhhCJH*UaJh8IhhCJH*aJ jh8IhhCJH*UaJ!jh8Ihh0JCJUaJh8IhhCJaJ,56778HIKL( ! צן|peVh%.h8I6CJ\]aJh%.h8ICJaJh%.h8I6CJaJh%.hhCJaJh%.CJaJhhCJaJhhh8I h8Ihhh8Ihh6CJaJ&jh8IhhCJH*UaJ jh8IhhCJH*UaJh8IhhCJaJ!jh8Ihh0JCJUaJh8IhhCJH*aJ !2"$$(B(,//n013 4B688;; $((a$gd%. $ & Fa$gd%.$a$gd%. $ ` a$gd%. ` gdh & F ^`gdh  !!!!D$E$$$"&#&F&H&&&''(B())********++++8+㑅qh㑅h93CJH*aJ&j|h93h[CJH*UaJh93h93CJH*aJ jh93h93CJH*UaJh%.h8ICJH*aJh%.h8Ih\CJaJh%.h8I6CJaJh%.h8ICJ\aJh%.h8I6CJ\]aJh%.h8ICJaJ!jh%.h8I0JCJUaJ$8+9+I+J+L+M+,,--..////n001 23'3 44778888C9D9E9F9999999999999;;;;;;;;;;ijijijħijħħħħěijħijijijďďďďħh%.h8ICJH*aJh%.h8ICJH*aJh%.h8I6CJaJ!jh%.h8I0JCJUaJh%.h8ICJaJ jh93h93CJH*UaJh93h93CJH*aJ&jh93h[CJH*UaJ7;;;;;;<<#<$<O<P<g<i<<<<<]>_>>>>>??i@j@A AWCsCCCkDlDDDDDF FGѾﳧ||||||||pph%.h8I5CJaJ!jh%.h8I0JCJUaJh%.h8ICJH*aJ jh%.h8ICJaJh%.h8ICJH*aJh%.h8ICJaJ$jvh%.h8I6CJEHUaJ#jg,H h%.h8ICJUVaJh%.h8I6CJaJ jh%.h8I6CJUaJ*;J=@WCsCG-G\IJKLNNOQQUVV:W;W\WeYZ$ ((` a$gd%. $((a$gd%.$a$gd%. $ ` a$gd%.$a$gd%.GG)G*G+G,GRGSGGG3H4HtHuHHHHHHHHHHHIIJK*K+KMMNNNN喅wffffwZh%.h8ICJH*aJ!jh%.h8I0JCJUaJ jah%.h8ICJaJ!jh%.h8ICJEHUaJ#jVH h%.h8ICJUVaJh%.h8I6CJaJ jfh%.h8ICJaJ!jh%.h8ICJEHUaJ#j UH h%.h8ICJUVaJh%.h8ICJaJjh%.h8ICJUaJ#NNZO[O\O^OdOeOOOQQQQQQQQRRVVV:W;W[W\WfWuWٺٺ噍{ocXJh%.h%.6CJ]aJh%.h%.CJaJh%.h%.5CJaJh%.h%.6CJaJ h%.h8Ih%.h8ICJaJh%.h8I5CJaJ(jh%.h%.CJUaJmHnHuh%.h8I6CJaJ!jh%.h8I0JCJUaJ jgh%.h8ICJaJh%.h8ICJH*aJh%.h8ICJaJ jgh%.h8ICJH*aJuWzWWWXXY/Y;YJYWYeYvYZ&ZZZ[[d\\j]k]__P`Q`aabb!c"cccFdedfdddddddddddeeϾ~ h%.h%.&j$ h(nh[CJH*UaJh(nh(nCJH*aJ jh(nh(nCJH*UaJh(nCJaJ!jh%.h%.0JCJUaJh%.h%.CJH*aJh%.h%.6CJaJh%.h%.6CJ]aJh%.h%.CJaJ1ZZu[d\\_e!foghii;jMjnqttttgd%. & F 0^`0gd%. ` gd%.$ & F L^`La$gd%.$a$gd%. $ ` a$gd%.$a$gd%.e!fiiiiiiiiii;jMjZj\j]j^jlIm$oo)p?@ABCD` & F 0^`0gd%.gdh$a$gd%. $ (` a$gd%.$a$gd%. $ ` a$gd%.zz[|\||||}}~~Ӂ6789:;<=D` ˆf&'()㴟㉅{jfh jh%.h%.UmHnHu h%.CJh%.hwX h%.hwX j h%.h%.CJUaJ(jh%.h%.CJUaJmHnHujh%.h%.CJUaJhCJaJh%.h%.6CJaJh%.h%.5CJaJh%.h%.CJaJ!jh%.h%.0JCJUaJ$`b&()I݊ފ<<bEST$a$gd.L $`a$gd(ngdBgdgdh$a$gd%. $ ` a$gd%.),5I1u/0JK[\^_܊݊ފ <XЎ;<aḃ̻xpxpepxp^Wx h93h h.Lhh.LhCJaJh.LCJaJh.Lh.LCJaJhBh(nh(nCJaJh(nCJH*aJ&j1 h(nh[CJH*UaJh(nh(nCJH*aJ jh(nh(nCJH*UaJh(nCJaJhBCJaJhBhBCJaJhBhB6CJaJhBh6CJaJDEHSTRSDE_`pqst պᥙՒuauuUuhBh.L6CJaJ&j h.Lh.LCJH*UaJh.Lh.LCJH*aJ jh.Lh.LCJH*UaJ h93hhBh(n6CJaJh.L!jh.Lh.L0JCJUaJh.LCJaJ h.Lh93hBh936CJaJhBh6CJaJ h.Lhh.Lh.LCJaJh.Lh.L6CJaJ!ԖeޗYn֘!Q֙יgd8I$a$gdo & F gd.L & Fgd.L$ <` a$gd.LgdB XZїӗ GIMOɘ˘DFęƙʙ̙֙יۙDIJhh8IOJQJh h8I^Jhjh8IU^Jhh[jh[U#h8Iho5CJOJQJ^JaJ#h8Ih8I5CJOJQJ^JaJh.L5CJOJQJ^JaJhoh.Lh.LCJH*aJh.Lh.LCJaJ2lX uRnt h^h`gd8I h7$8$H$^h`gdh h^h`gdh h^h`gdh h^h`gdhDSlm˚&JXYޜft !N\uvϟRS^捁ssssshh8I6CJaJhhh8ICJaJh jhh8ICJUaJh jhh8IOJQJUh"hh8I6CJOJQJ^Jhhh8ICJOJQJ^Jh(jhh8ICJOJQJU^Jhhh8IOJQJhhh8I6OJQJh-no֡P[tuѢҢ#$q 0BC ~)*j}§n÷÷÷÷hh8I6CJaJhhh8ICJaJh jhh8ICJUaJhhh8I6OJQJh jhh8IOJQJUhhh8IOJQJh?tѢ#B)rԪƫ9nh7$8$H$^h`gd8I h^h`gd8I h^h`gd8I h^h`gd8I h7$8$H$^h`gd8ITcrsԪժOƫǫɫ189:]noíϭIl۸sf^Wsss hh8Ihh8Ihjhh8IUhhh8I6OJQJhhh8IOJQJh jhh8IOJQJUhhh8I6CJaJhhh8ICJaJh jhh8ICJUaJh"hh8I6CJOJQJ^Jhhh8ICJOJQJ^Jh(jhh8ICJOJQJU^Jh"nQh r aSh7$8$H$^h`gd%. h^h`gd%. h^h`gd%. h^h`gd8I h^h`gd8Ih7$8$H$^h`gd8IϮ߮(*QRpQtŰհZrӱGNhi 㴨㴨xc(jhh8ICJOJQJU^Jhhh8IOJQJhmH sH "hh8I6OJQJhmH sH hh8I6CJaJhhh8ICJaJh jhh8ICJUaJhhh8IOJQJhmHsHhh8I6OJQJhhh8IOJQJh jhh8IOJQJUh% EHVrs x<Pab۴۴ۣt_Ohh%.CJOJQJ^Jh(jhh%.CJOJQJU^Jhhh(nOJQJhh(nOJQJhhh%.6OJQJhhh%.OJQJh jhh%.OJQJUh(jhh8ICJOJQJU^Jh"hh8I6CJOJQJ^Jhhh8ICJOJQJ^Jh'hh8ICJOJQJ^JhmH sH ζ3@STHIǸȸM{պpdVdKGh[h.Lh.LCJaJh.Lh.L6CJaJhh.Lh.LCJaJh jh.Lh.LCJUaJhhh%.OJQJh jhh%.OJQJUh(jhh%.CJOJQJU^Jhhh%.6CJaJhhh%.CJaJh jhh%.CJUaJhhh%.CJOJQJ^Jh"hh%.6CJOJQJ^JhSHǸ KLgd8Ih7$8$H$^h`gd.L h^h`gd%.$h^h`a$gd%. h^h`gd%.h7$8$H$^h`gd%. GJKL@ABCETUVWX`a￱贈pbh%.CJOJQJ^JaJ#hq1h%.5CJOJQJ^JaJ h=h%.jˏh=h%.U hEpjh%.CJOJQJ^JaJh%.j+ h%.Uh8ICJOJQJ^JaJ hxh8ICJOJQJ^JaJh8ICJOJQJ^JaJ h;h8ICJOJQJ^JaJh8Ihxh8IOJQJ^J""ABDETUWX$a$gd%.gd%.GC$Eƀtgd8I L^`Lgd8Igd8Iabý',HIKVWfh{оѾ#$6klmnoݻݭݭݭݭݭݭݭݭݭݭݭݭ݆t#h8Iho5CJOJQJ^JaJh8Ih%.6CJOJQJ^JaJ h%.aJhBCJOJQJ^JaJh%.CJOJQJ^JaJh%.5CJOJQJ^JaJh%.OJQJh=h%.OJQJh%.hfDtCJOJQJ^JaJ hq1h%.CJOJQJ^JaJ-()*+,?HIJKVWfgh{|}~$a$gd%.gd%.оѾ#$6AV & F hLEƀOkf.^`Lgd%.$a$gd%. AVgQV & F hLEƀOkf.^`Lgd%.V & F hLEƀOkf.^`Lgd%.gklmnogd8Igd%.V & F hLEƀOkf.^`Lgd%.21h:p8I/ =!"#$% 9 01h:p8IBP/ =!"#$% 5 01h:p8I/ =!"#$% 9 01h:p8IBP/ =!"#$% 5 01h:p8I/ =!"#$% 9 01h:p8IBP/ =!"#$% 5 01h:p8I/ =!"#$% 9 01h:p8IBP/ =!"#$% 5 01h:p8I/ =!"#$% $$If!vh55#v:Vl t0.'65}DyK _Ref146371092}DyK _Ref146371176}DyK _Ref146371411}DyK _Ref146378232}DyK _Ref149220859}DyK _Ref149220900Dd 8 b  c $A? ?3"`?2h7/V$ߦr0&@ `!h7/V$ߦr0&@xtoxcdd``nbd``baV d,FYzP1n:L,56~) @ sC0&dT204t/ @2@penR~ W삏B ;|.? _˧ĤUsi#.[&=֫2(8n@L~L #l1 M-VK-WMNPs_CdB^%&0?I;Bq`dEyc/(hj@PXWRWWFgfӈH傦.p;Z=TĤ\Y\ rP"CXY=t͎Dd b  c $A? ?3"`?2 h\= `! h\Z  &hxڕRKA~֟J""dXAt6HO6X)jUPy_ ѭ7"QAۼAM }73A@{d|%EC 1lP?BA+mx y)gAy{)̠`چ+tol.LBH.l>铪mV~|6;ÄμafDi9/E\H8pI^/(mD>B׵am<~@r.$]E>^yO|G .3Ju˪vH)0c]f{vJto u1b @J@XmZt ˆ!7" wD/ڕ}Dd @b  c $A? ?3"`?2 Κ@ㅍ7B `! Κ@ㅍ7Bb `p:rxcdd`` @c112BYL%bpu 1X +|-(T T~03>yN\q?]d6\0sA,@sC2sSRspb.#\>`fVB LIL>H?< *a6łb񳱅+_WM=ׁ&KJ|`b%45rS,;д(qF&&\ @ ] @Yݏ`w}DyK _Ref146378610Dd ;D  3 @@"?}DyK _Ref149228353}DyK _Ref146378454Dd("%VU>  C Ab+ɚ0V+v(󖊙o n+ɚ0V+v(󖊙PNG  IHDRGKsRGB pHYs+IDATx^xWg-nXp]ŝš_P NR܊{xYe]M<h =w۶@KzD^a֭aüӦ3D 8<'<\6p`|:jSx͙iST,:ˡ Ѐٳ&" H]:dPj@~Շ{Æ~ehBڍ}""d,_W&NqH'zYZs?se[{I;I,AZ ,+C΍Mt {^{S'>%$$L>bŊӧOז3g͛۷/T o9  !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~~Root Entry F`PdData WordDocumentObjectPool 4`Pd_1219177575F44Ole CompObjfObjInfo  " FMicrosoft Equation 3.0 DS Equation Equation.39q‚0vD  p(X i |parents(X i )) i=1n "Equation Native _1219843340 F%7%7Ole CompObj f FMicrosoft Equation 3.0 DS Equation Equation.39qX@$ k(x,y)=(x),(y))#*# FMicrosoft Equation 3.0 DS Equation Equation.39qObjInfo Equation Native  t_1219843786F%7%7Ole  CompObjfObjInfoEquation Native y1Table+]H,n  i k(x i ,x)+b i "Oh+'0   @L l x (Semantic and Machine Learning Methods Biomedical Informatics+jL̗O;mgW%Z@YKG,,C80tYuo;vX||R,\p?ŋQ'uqaÆ'MfϞbŊW򻸗B`_/^#`{UB [7_|Sc&Px߾=l\3sUV? 3d 2,Y? urҰwc ލH-0v7uE^=K<2exyyNڿ/“juiT\G;w,ԩzʕÇ_xQVIy-W\ƍK*eaĴ(`aq׬Y3dȐ/^߬6 @.3OtttDDijgϲgώ?<6;\4 P-v,>ZG6 ~z}F*GܩSl^|i=}۷qh|k׮5;q_euEW?~\n݂ ~GεkΞ=uVdF6m6m ]z-[dɒ%cƌ߿M* zx, tKHYA._lJ.+W,ZUC؁(|o㏍uݡM[ٳUV#eu<~҂f̘…#'L F01118Stpx0!g ɾfiB.1@رNj;frTYfQa Ҝ(,/.PATR9dGVW$D[t)<㬐n i֭駟8?fW&sN:SLίɓ'4MI@Nӧ1”)Sj֬Q- CI$ |- Ae2; t5i져nH.[}Ç{,c+fkn_3@W!iASnx I -HހcL'7 Fu0^駔Xo]v8\>iC HDPbϧ~"?~3BѲj߾=3cD 6l6<3›!;y[W0^3%Ga,,O~،Bgf3Ȕqf1/tjB ܹ#9k֯Ju萐:S 6]0)Α3D\5 wll޼ 9[gbGX.IrqU)h-g#flLddL3 lGobŊt.PM+$ |#C?c"PG3iv3 іC le .tԚMi :?KW 9阍!.JΎk6TXh ].th$ʭ[#Ǎ~Q 4!0>NoO2s RJ5EWVΎ) J]mM("vfsƍypO%rK'Y0M{5*n_7XeD?<6!>< RSvgx!"ֺuk&<@W+b_???/V3$g&~OIsP4fD#t8i- H&_ ;fpߗI[KB mg?PYKI$ c ,Z qw bR2 I~wؒw[PM$>8`ZBbu@lG hmRGY4Rg HO7v?i? "vgi/3 H&G в9,sNܣMq,xb)7NkrxVgSF` p /P"qfka9r" !f+Fԑ"ڏ7hߤD?n+F1, `qCZEB䲤4d@L ~xl۶\g'*Acי0 F^I47o ɐrЦc4 Gi9h rG¤$r#QmXf%sma lDR(N7FK&w \R?K1.q ~aZZxSOx@YYIu wަhlF?N-sh `H -H`bЄ i'3iv@xP֏ (F,Jv:upU|5&HxGRn czzZ4M;V}JaN8aT@7w*9$#$B5y?")wi `_Ɠl LfB6 #hyM6Eo/a zcdŝzYچ VN:ZX8ltܸq4 #ﯠ~bL#8Ud;F3n"r $hLb2/I]iRG݌C!i@,Pj&$IPgϞxc<)&gGX~=t_9.rMO֗-Yz9yQWؠjZe^ X> C?8X\֜9r9J'N9$ov0v2~xoӦM*U*V CrƍC |\1ri>;㈫d@ cAB[Ҳ8(L+a4LgfĬeqj1idd$~Cz<%{7/j1cd qÆmW{ =p`@"AIT!'D ,ٳgPh|CjHի*쉽ၻ^֠w}'F_T)tE r$^⪒s~UH3 H&1 Pƣ,<< 44jN] ը6n5[df,v&> %= &G<ܡa˕+'rE:xyA0U}ok8:)@F~`*Aq] E7 H}!B*2&E$2\ؗ/ n}WOZlC(Fh=nر+D"g>75,[7A*<Xz6x;d( O4)6RHA83 f6cɧ<R(3&yqXئ:ua$CWi,(1+>0 ҈ITCĬhүa-_9w@YߧMuĵFDːQp˔Q1'*Ib9CxRKbmCf#wOa4˹|:<Sq)?v|J\=N1S#G~qk]Lß^%;{Y85q[t+qJ#/ 0d$Y#y2 T=H̔)>ox@ѿeKN}V tiȑ: ,{U;'`Ę~4t uRbTl9T`8%\Gڳ@P=v"@?n`Kt_+AKgf7U.g/c8$}iJ#f6@@j1id$Y"ePⷔ-f^iZ#vQqs_WOSMZ2*O4Hj|?|}{*ꈿHʶm Nf 3<@¡xĘ$t$} Gx#I xfpA8|٧ :z,}BfI 64qMʐHi$HWԜ(@N";y=lbYu& \ԶmыP PW[Xf!܈#y¹ٷ3(ʞ=;ґŋOLy@P7*sK2Ę$qQVl_b2 Q Ȗ+;Ba w͡#sV)l:c'kfbufWDؕ[q֕|5% ,$ ߯=VdɥկM-Ɣ?^]~`1(//U鞋K_ ƴ*քCԼyFEOL)pdUlWNϕЋ1~I>}b"}!\'_b*Vb9 H\!0)cJw@Udqg)*pź鱒ɿK ze iD&6k>v[9s|zV\ЩԩEꂃϴISS %30Tw۵޷ Os|T?DGz84&2y+(tI!~Rظ(6bC-'i͖itLP\%HZSxέ3m:!q3AiJ~k7 F6ucS׬V=KԺ),L "8X%KcԩSqr=[vM''˜G뼟P`'Mh*sl<qBpڶb~[HgJ!ۊRAH籺|;\4h դA;3Fz0mjXoz%t;cY4x'z1lhQY[`֭k?ȯuk/&, #/^T _4 aG11oΟ3#[$ؙlK2Uop:w9>*K#Wrx˳degdd$%}$v9;k%>i #6 x@%Y7bbfm˖-S:MƝtv nL=lݪBX&;u4{U6EUׯOQ=;\A _Ӧ-8,]:`*U˔|=Z$\+(k8rt"(LR+w;FxHׯjU_K4kCRsvcl$yf*<ȷ~ =U&nq cHIQv;tZ5`@*-vmΜT1B`V\ѿ`~Ǡ)^Iݻw-y[Ld<}+Sj6q2ǵuqke<?G6"|ԖHLk9yIuuiRM_q<|\,HW`YDzp?,Ax~}1b@@-O9vÆ~p4(:݃=dHԩPE)$`kwǕԿ2y@bz;5efO|TWt 91OqEC|Dd[<|o#}Eo=T@hтvnb{EaƥKvЁ{'P\]ڝ"e"*'li!! A8[f~:rÇbXyk1֥؂>@βQ3b$nD.4$m[@qVH"`R,Yi)//PA $kє(LAӷjS -cF)HRpϞEZun^țE(~]~mF,)3~^ᴶ),v 3Sq1}T -b,iiM ^e52I"Mmtm#Zmk$ϕI<1$|}toǀ5\h;.{ 3/ߊM[(8qÞioQW}\9Ç 2f`腲eKEibv۹s7ӣG";^r%Bpϖ'^ydu5:vF$0v nK |)AUgM| deЉaPjM]%_D0ϋ5SO=/ڴ -zz3̒Y;&TRԷ #rVH@9}:}zLݗ_hQ[?:y y5]IjŲQ CQLB奛5NDBKx,h…{0rwI@ҥK%TX jո}ֵ3g @i6h 1E4S/ش K:-Y1i3QS.M}ԉ5jX{9+x?\Y]^B[,Z8Ϡ#F\oK]ڷMNfs 9`U-6Is/JÝ@\F<Y|fOsZ{Mֺy-O6 |} LĞm ܸ?|5WMfH% ,%gvNjUUNuAx{$V}b9e6rBPɉ1cgM%{By I-[7.|JMy*ӣi=ZYLҽSvi/:RK+HdŅ >c=:YRϟ~'I”Q =I`%FF >> n(.J8#} Q't!!TGe`{{kK?_eܾ ^x%/K@9:iԂ"]x]>џN[$iycKzi?@4ҠVK.mxҩ|B#tJri{@Rn6=x`̙_|C7j߾<5((8gܹƝ}.qRN?^t锓j|)xQٚ&]* :,.r'dP|^S#ŵk) P{+Yke"D EqM*~KL#zM=f KMZW%ءLV`LN/=G7˕6_[zik\M5}{&/- ?}!HѣA#G/_>@,(lN]$~ Y2D#zZHS*| EcfU;r>f⇤`q;slRPUI% ':wFʪ-{ >gXMSX#?ѕlhT׎f:x/[cS}ʇJ)4.VZ~ۨ~&6hnؐ q:vTnެ{he9=;^בO8aa^Q'o9'd Y{_'d>6<3d@]iL;g.`@&v:RLW3?k֬d(zTx`8jEŌTf&> v?XTK)VY'ܶ$T{/8RA3yS,>3 4"捸vM7[ 3Sݻ`oVכZ s~Ȁ\T%)邍#>EB(UK~l3>|hY޾@X\=hY\/x_sD ;ʐC׆ztDyLBF?icD4ѣƃ~7jڼQ3n믂CnԒs bNf̘a ܹ3ȚжIj5 Hl%,@))>׸qqÆ<_ |BJu8l*ϦM_0l-?ehX L +xq Fu:D8lx-3MKf(YǏ.׶- EMZ@ncWc-%HiU}\)eߊ_4?VØH]:&_|&mU?E\C'. QBƀIhmȰ/">i݋>{%T&dQ[bPq$2qh~z˻vFJ*K %{IFө)̙%#  C+>ɍDցjGmaϞRADrRgzd./@ ?"Ha!Y%3i /Y- fycKn4oc'^zy\ʕ5+«- 27~).]N^^ߍL^x۷oG=odbҤr>@ Ӓ>4&PF\&7ie@O>~lPȔqt;tVo`_zOEqvc9Kf4sY<珲hMd*](ce<i\GL}L]SΥbYtB˖~ hP\2zӪ=dz@rLSNyv.YC94pe%dh ޣ!-ò@uVPmdKOۺX`7wx5HS7j-^ܧ[~$zƿNTK˚B m !kVϬ+w*5wjVaTRmhǢPhsl5[_?/3g~ĉKz|~>w&\h'nԂ;MLpC7DF őBg/T%f}5B??~h"';lY5ZIXk>.$ ަv1C"nBW/7hoqi(Al2eS-_ޘ4d|[P-{1»H@&=W.PK4~Qey>*w'&U ٴ_i.]66TW 488'k*5vc2@?{,6n=v>FRzl0]4 $;l#H#CHq",/Gb .t PAȑXTKX䅇]L״iCG5Y74_TiײA6u

Usߟe3k:*m6&N=*fS6de_d% &V-T޽{oAmNY.9l@ںufϞ}yȓ,զu# P/! 8scX<\&%Y`m[ʟ6f S57W$ɸVԢ.{D:[DzC]Cۃ "ˑswb2gVv@:5M>":įs̋{^],Z-f]tN H۲e -֌@*G傼Ű)zGں[n`[{'g7|pg9j׮toE ; G۶~H{v$tJ(:8P#y~^5cǍ,Ac"iH)n3klcu5z)k3Ca$>^)?5Ժ6l0D# k +|e1mQNe^JᩐWo`W5!}H^p/޸ K_͖'TOiHe:$G4SU*N! \ӣ :$.Ν1I1Og:&X$T_\ީX_|]:\N66^P|fhOr%7={=_`]5{y>}TX1 ӻQc3@%YW3dNY8$e"S:$_@Kzhdp# dG22-ͬ|B!h-!͐X"hF 'jQc41w #2_AQzVxFU*qhz8ݥAAY""+װz]|8mI8P-v,dejFl_2'=?&u:}E1cbEsFUR{)Jж/sη[7]@.o^U .x*JoU|~yT!;u*zFH[,5^mi ~ۣGWV\z}M. G5ЭQzxZ7A){kH \ꁂ;Eiߵ9sЧ_$]FZj@Qhe&o$UqTGnޔR עmTHIX_b$zҾfb2 ߆n1$ҥDkvPԴec2j\(RڇF[6D/ȷR3,LusC7&c~^ko m-t~N DLT>p<$.-Q,:㣉ӧ 2^) m2Tc ],^+H)z=uIeyރ2. trV5Dr6U֞={6[l4j !vc׃=B"7We Ql#,YPd!Ah,5c}AW"F_sD 5R:DU(?n݊5/W.2dĎ,YL)StiJXbI-^%vx.;[2]SZ35s&p{039rPY䫚Mի.!GoY|Q&]' AH YgNk_edI E `ѓ̴t844,vxΡK"<=;|}}?6o=j++?fЅOޤ~P㲗T^oM%" w&%M@͖;sMcAw/I]'xfR͈`,,͓߾69fΉgTdCѷ_œ'/?GąbTB"+L.9|'HND3gN`v^yjD Մ{Մ/l@5ϓqݭgڵ)\8`9#q#p4$T o8qF!0'03v TDC$1-򪮛=h~e֥N0 HGÇ[~DM"~}Xb%K<>풭][f\ȝ{dã&t]1)4# /f! y|y苐ݰ{9l~V|[˾]#4"XhJy */W|S dժU!j `ApjhM$@8a#;Ki 4dHJ 5T<\h&Re8YgA7퀼BͨU0MgvN-# ~4EENg]Cjn-D?BPX\b}<"[3p eL4ҿv"E(YK^8zRzD\fh^xH(2YJwↇ;)SfGT[>q'1^%)+2bvaAzj}B!;n>>E /y囓k >%} :jtrE:w&ǣZi]-km?\'*f)#gƳEr&sڎ2M+>$JӢn7oo?tοf vF3ݻXV ̝ ZPN/?ݫ޾};AH'K'>%.O0)F}1Q` 7B=R/C؀j++lN)5{ͶWܼ߳'|abw^WȣG5ի3XD;Wౝ%~_~Wv>"I1'$$ 2Ďؾyʱ,?;wD9\:~]v|bw&s7LTo-@qb#簰4a_v0DBPA jEb-|o+o AAXF"-B cE=-Q cݺK=*~~t_=iϏFf.zď?6HuT;ȷk垥Սo{Q6 MGYظǴFhSA6l)+ yFz_*;~8d?Fׯ:;Yqqq̑#0$+TTӧQ-^Wfꄖ-z{2M?N0,L TrWrg7}iWj/AED\e@;0ުTWXf͚-^h/ r!z)P@۹kB]ppG1}Aъܽ*2`][[qٱb5ujcg86m T)`Ψ|!XtcS;~ L@n<ŞХK~])SԊڶ gM8Λ8!CBq4_}A$@>51cƌO>͒% ߿#vAq ƒ}E5^PZ8RL8~Jx;Mu&N^-[rjӧL+_O&\ғؕ 9r'@J=u@x ?i 0õqޘKcԀ%,ZQD o۶jUmtW?.R$pݺlDF/Xp֨FZ4 )5nv Vm?v&(IF*KIqh'M2E1#ltyzbօ( 4lSܻs*l($ԫ)nIpKC/ѣG92J7ri-4fs q43f }$\x$(Խ(/S^r UK!lcg6,<z=Ν[lق/%\|vzC<6 kM0 ե/CKj]KlIdd)^ϟľt޹S%D+xV/D_39+=HƏsܘŌKD[rg58~;(~1H͛es+o{رH>~y~?ٱ~mT ^[3f;FK:Oԥ$=2ր!E \C83,X~j.?Y4^;'›VA۝%˅'(54ݻwo޼y>) j.d{PL+_I.mnIzJ%3i ږ+Ҹqbpp.^QgҲWK4$zHRnEJ&&[j*q^k1c]6f=*$5-e"$5kBJ'f4wwo ov첣y66coVޓ' [)[l_f)Οwpٳgi[=z4C-u=TV 0@ڹK""Ư @2F䟚7opѼ>;}mdS6EΗ/Tؗ t:E{W3݈FBt΂ee!)6mD6ΡR ï ~1S ~̲{ane֩ri8gnXfpY=SЅ*74 YfTe1B󉎖ݺW^[ߥ'H6#ٻw'Ofʔ (EYqW~ew]$_͡\@{_(i\yEK=7#i \Q^Fl۳qiO~seK@;qc]ՏH7)0]~a[aFUexkG:nczد@ߔ-)CLA… 2j*cٍ`)2Rv\(z@_t8}X?L1`02YBUju:ŋ]fCFNBBOu0xp'ӧO>vĊ )*⨶DO6 ~**A+vP9;Qplok_rT 8ȑ#jqjz[:VN^&kF`[[W)SC_]ŀnR ڋ1rl~Ԧ@eqR zu*Bb W2VjY㑁}IyE;c`\Wz1tPZ9s>BN}z+(F[5ZjK3 `kۺJ՛b=y!sfd(׌;=O.J֌|GK;v.ھaC+<5eӵaSYMvUBewܺdɩ5k9wڊǗ)dI(8ZbPa!=u ÇIߎ>W85=8'Rإ90`Ymk&0X93fDGte$4l;)w< C_5V 0/tx]Ͱaq؎S~DU hf5 1r޶kVۗ9j%n^Ib*k ~枫a_ڵACw(B\DT04q:v9=坑M=Siv-^ n?m{ށDeD;DKQ:5j^^V?0OG;igN})KYfbܿ 2O+~i}S" \^taAuZ"W[!/bٳf:6 I&EWR%T,dۛ0#̝Ko-`aW[;:.d"#tY5C߭Æ0SOK<\?E-b knTM lp)'/y1M՞=]{fݺ VBGٴm\i;=-톅(lY%b&L1ʃ_tLz7N?]<Tm]>VJ)#|!ƣQjđ !W;%Gkw#=m{ʬE|56\A2ߗi& .[HMM22W.k[lSNQ 96#IqiwKʫoRDI2W5As- G40*⷗ Ke3JYn=(+-|!&9^ ό_x`|X͆Y+"R>,t! :t`7,"9'N`Q|P{,\(~;? bv1Cv]wJ]4M4'i%u:@0ݍG&N.C&{H._m+ /UC mL}/D},ZQ;MZMm޼@$3qbΝ//^ܳkIit5"fbKH{uq+ٸq##GڵKC$xPki.]Yw}5 ^8#NM-9mם4Pw}g*s?'?89&fuPGQP-@ f>;v 6^dV ~ ۶5~1% 5(9CZ]&Wde5ڴvO* g4tОEm~r$˖}lAãAC )ZT {BlO_o 2 sØ4v-APJUFȫ8pܸɫd!BJ'oo4ABegL z _||yHFNQ_ CҰ)0ABw=ݛA\l`b1I5.M( ~@j /|)Q_dBVmj1nܹqCklV}sMRaadi<;9{6Ec!츘9p1z& *kuw{JRK0=kbhU<ꄶHGj,v b,[}+ᥙ:X.WVk4Ξbʇ$ *48xMyXGt+r0(H_)R*u0z}Z)AC%DoLf1x+{V滋QKJ_[۰} Ь^m=_\i5QOnǎABLkn Iau^>H- UWcK2ܮhni=w솸s )"22($O_̵~,BIO;WbLʳL Ȭ]PM&:uo}D@jm<{Fo&<نme2{׋+X'fx7)"EٷjYnx4HJ87^N#T1z[ùq ignN"FKkUm|y0 &f39NL@{w tg^u}h{.WXjFRqsw&ԝ8 i6nqY޻\}e+!0-kĮfԌ᫁3PCs W?Ӿxy u+O^OBjl 6knҰ6T3sˈ1$< QzfZfay9/d<X]0oͼ2Ԋ ّzRIp5'(E40[ ~WFx3hSdQ1P-VLFޛ.Ѡvt(.^6w=8[_T~A/RD5JkՒ_8JVP1EM'ڶ־){#e:2W[JӂPȒ7Q3QQQ ]6m4`[l)k_{!~KfCJ=ƗP-)w 2T(jsѧOwL"YVsnTT!fyg+!`+LdȮ%tdgLHBV~f#biL"¡S7thuçorsOզIi\mP(PҥRLZuLe V6tkR1,CjP4QE+hUѾ6Kh];OMx=?3B› ȢښީcCk鈥7]ҽk; XidqeHg3K#qGmRYQ'(ˢŜ?^ӃlI*-"r`FaCYА]kR`]˴dzj#$W.ݵkYY+Fm*'8\᭺((pgpJi v^ sBZ _Æ[s6ϵ#d#~_ILb_Q \S/"BvR6Y̴GK&vÀ}}ycjO?YtP( c&D ;qթcp>`fUj~ԩ hˉgs4i~~~Cr ʲ*wL:H9OƉjN!GgG;wvOO] uAߺ5G ?T4bWD&{)Q7B9XgqcVO\1Woqzݔ;;@cɻNo)FM)S5wF;Rpy:x{oiƃE-̣g_ܹsחƶ f*=sf+-ƽ}fC)N f[y'6k?FIO4i PM/:wR`q4{vkPT5Pr(h&4g]v3ka<>0T 錂b@n֬JQt1 DwF*]}I}/YĖPsBua/>$p9/C-]k'ȢuO m:=4ojxc0iA!ټYP-:{@ϸpuj}#F1}?Fމg4pe#{% G6o~BJ^NVQqrfB*w雒K3Hݜ_HH ߅@Ӛ'HzI4Ȩ Ido! Hnܸ?<8DĐtto{<!9cFA<)ƁXM٩?Ǡv4>Se|Ш  |5c!quv;l׸qHT;֨ԦU ԯ@uI]iD"#Y(EҞXhg>#mI; |H֥K գ$8E}R" bd߾}-]jЈT yǎ~-WkT6WOJ6,zŀg)SψCvdoC[v钲bE* hJWƉD6J/VJu,*JӐ4P7hУCɼ=rL?SEr-k)R@tZHٿmA j6-mQ0 -ƍs`PtۖI([ZMߺuk^D4?/_ޣG>t4~嬽P{q"/T#-أ3}[`#~}H'e,WhV|Ҳ͒lqQȔ]2LڌNֽz6k{U+ujmg@;92pԷlfTѩz[S;6]vBTH ?fT僃99?Xwd٦WxMѣGLwؿm|i9υ =wfGi̿qgM A&g㐖xE05ܐbG\U4eCR xiӦW;ٳ'"秒FU @ң_ٲMTi/XEd2m-Y1fL[le3JȵG~"+2G l/s&V;wũSe7~$o~}ue[=%|C cnj;vlŊuC:dm;3.kVJƔy}#IT`p҇3ܸ]) 6.풌o%tC" 9SW΋- [,:@Ί7o"qqy&}#*i#G{wD=}*\Y??%||c\j JebQJFG2UȱAW,eڜI뵑#/6̇1ޔ!Gx,ϰNƌ굲pݴ)k{"Ub"HˡlgF2g(I;}y"ȅ9rC:Jza9\qߚ=(eƃW-.R?<7U=ō;g3S?x mjC>>/(oax.Z.4{boUeҥKCy.;(D//nL͗6QkPP?      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~I>п!` X? dxzg  iB },f3 ]FSY3IFxRx.Oof+* ?4T5ڊΘ^VzOk ?X[李1hI dUi2r2օq.}PlN\ b{;ի+Ο߿o52{kwW2Td#gDa hYI/|"8 _9#=v{-D o ( & 0g[%7Z=A|JB gC/_KNȒXb"vG=~SHchΜn~#tO}>>S+69:tE"h&"SP,w+$Bh{ lC+U ̝X(s|}G9<{v4iܹ7 I w^Ɉ&rtMjfaqsyaԑC.0R:5cv]z}pǪU_|Om<ȌНgO.ZA6V F DqhP_!hf:TpFj>%By=F&#s6B۷_t)~@GԪXƒ@B!QU0ڂ4¤Δ6#i~Jr/EjQ;CZ^bwh ALu<wrYjWo_h! E^0K=dcF*+ VDF+{ʈh &E58 fMt\\/{ k 2]:w".y.]:;hH?gwـtǏj+ƜTs[o\:9sq2R%G8[7 \vwj+V0XgsHw؈"{o:3czR[h0 9B%d֬Y5 O8 qQ>KyAJިQ_*^+< ׀XRōAbv%ȫj6`: ΙWY=4 is{Z\ex/c4_IP>,Y&^IP1}|><=}mW콷WGq!?NYVːA-[_GG '>^!˟_hCw!eP"h̙״iBZT&>jwix\Iٲd"E cu8Aōʕ};uRށ+ Ӽ7bq fW`yP :qN[:8`Ho`!S̋şCO?5޽{ 1^FtpFHl/b| y:yMk4o]kg/RtfǕ'3O:t~~%.vT0'Od ]zp$)oY~+ uT/Uj9{F!WWmNHkB$GLja͚'O,[װaB֞s&֨lYo ;_ 04LQUHcG{/t˧?M<񦝟/).]^i .maGOy`'DY6A 8]aaatD5TPɶѣG$VZڂ4! GSuf~Ve)PIL]~IzY4A?^[ilkh<`xc'^ƇvM?wl\G $441!F;fR)_>Rx]StG&1EGon޼ߗy<]n煍&j4f̀=gF rHuMMGҭR2%z"8HhKzxG[L9QD o"tWB. :O<2 ,At&AtQm۶ ▸V8aϠ]X ru|**#G+zU| 29;72MƷ3f͍ҽ]蔆6}OxXAOOIqO ?mX2psl+W>L ʕLأcDOPW֭Sc&o#!!"a<.a ʞigZ-f5Zzvz~WO;=bDZ:&T˸z ^"yW3f7쉝:*^ & 3}:umA]=4@"V\95ORzn8x!%0Q 83Q.B Bx=$@3Ewo۷oUkx`OӝYcSICQ62sL}Y<7ZjH5'0@ <~<2C+.aàekqPT:&QZ&G>V{&N5:K6kѢZ1S4YK!M"hx{ ӦU9Ыj^X~:Ƥ[G5Bv;Ha(<3Ss!nЬ<@ F%<~'Wb5Ae=PL Ⲙ'0XI‰41yfy^DeJk~!&DVZ| )ُD{YOz* aIRIh^ܸwJLg31iN :zyEDw<']Ǯ*jG-ru=B9s$Z5 g5'N" flócN&U= y_%ţSIy,Y;irv3_>ٳgPT'm:A2n'Hh/?ܰ3#OH==e *cz pnPct(e^ڵsrϞtt(6Gj_+UbeM/'>0X2hbgbrHPPBeњ:_%.9ovgzm=yRH7m&-3J ˔^ no I̟_q#h1s gx1=q#=Qc\K.A0':w8=b-3+@Ptg/cy|ĖDWwjϞ=WKwXj\j.G ejޓ'Em(XA 釚Pj@,mۛ\޲lC]03bn IPMن QQ``SXBKWAd`>'{ 7_"$UJyj0^"A'hK3]B+UuWk $5f YܠA$c,PڕX`Qܳ`jIJ٢ի߈Z޽ vYڈw1rTwQ Il%`3#>2nt |WLwxo)Z^fJp$hi .l9j.̞:5)_nt1Z7hݻ4I}`08] @-T MU,P'`SN&5P+#lnN_}UCb.]xh e+Z(:2%ȖY2*0Z@>YMa!M!K EY}W Cnժ}

N1ORL*t!j#N&ͦ)ٺL+& ЋDV&<|ϖP׮)w/qbͅXCu ~ŗ/CHxǏy *m0&tNܓߣmS"SG=zDϞM-|Rl%-VPM+?E~'jxsF >Q=>3VQ+Wx& `fM'nB [lg#ok3fx!ge6s|Źs8 ƿ_{ gFYUSʴii4 e2Tɋ\V(F=Ӊ! "Xxܐ!h͛fvZ5&, N~l'iUs.]ºft2xMM4K5h-V GDy);7oT.wJU" caՙ-T)Gd9/rn2, ZLjթIڛmob3Ncv8$hQ=VNpb€/2GH7.rza!X.\G4Z>yw~$E $f0 רA(%:d:B8p㌅1I@2<k#B?c(QF<GBaI 54jbMٮ2*!+Rx`MT#eWheMw>;>h!n/`Yؔ5+F+]Ľ7݈В%r+WZGv)7+nH`W~>n=j&ezdI޶`<3cS$4}qñN#>eHh x@;8y^Ͽre -o,~+uvNJ-o>uM]UM[M@GLjo#c< :~k|ܲNyxO?>}c_,H5y8b ӧǘTA hǻ#;#lv^4_av恪0(4$Z0%كkimLoU%'Q;BԦ֢5)aUF% 08ǏGJƃ iD?(6}=;rCFۋt+_(rWUҎg .gWgO({-;O +ktv"?NA_뤗[?&sMBgD 0]aMWy]A_)%o@njjIt#/E޵/J<!3p(!,WO(I:wBhQ]PFB-Phd2C椁h4tN?r؜>C@dSGtMyVXJ\m^Ou -M -7TJCn*}f  54"~rT2/ezp!Uz!l27rˮ~GrGӊ焂qM\k! b[SxAt9Bm ҡ7SXrԬUx*+!ȋPMԑ WGt93./&֥Y5kKy520! K"B7wCf wFeL}>8go-}稗Vp[WK0p\HUAKu=yiG$"uJɟǜ%?{.mܭ<цPK'>p+ -Bc` rSM[ٿHAX4/Ho tEB';%JMOIV< !;Wϟ'{Q#F0Sp>x C>|8$B Akĺca6ħL=z>/uYUU:!. Z'oyOFR2gHя+8%Ce#0~Fzuxj;sav|;l(SLij%ً8?hTbpC)^aaX. άMV->ه4Y'q-jgx">NL'dy"Cnt 0fT54IqzLD 0_7W¥EBi@Q2<0V)p~N TIx9SRif="X~n̰a3S_/ t ݶ-*K7oW\$5B'ibhO]uko(rV<~q^ueESM"MqrBǎNʙ3H ! ;]ϿZU&o^J3k{s_Cc6oseI J|S-T(iSCsɗ G@[l!Jʨ7o (:gy"btAAayݙ3wuj7Q3 4tE"5>n itAzg|5gMI4tqQh3W(v]9eeWkT7Ӛ NF3Imһ)NlHt#@!ޥG eŔ[xH@S,Cpa1#.wnW-[i?$B({wDT­ڿI`lBbwΨ#HR r*( z\QQ2([/!et Ui?7S ɞJQ dpPV*v'֨Y: W:6bbd^B?*:}4z}~}FpauI&,UP$kh'C %>6RafFb=rJC ~/ؙY*mtA= xpW3Kh"3V*Mtߝayi :4(%õL5 >⩇(ξX0'T]`< %D `7`)Gr9 >#C믿䟜9m>%tfpdA9TOС[偧/=b5O2vK$Sh.vmj]dK1+**N˔IGȑqG(sn,TNJ<a0 =fӤ s֡ua?ŊQ| <êU)\/ldj>L1AK0XSQ|B0o&6z.DDTH*sr|<hg+382n5#0? b9B[IL w%gHYSvƐƼ% juifrLdy KJD)p&$Hn@1ƒLw"U<4"ˇ{:rHS!;aR_K7u[np˝PY/@ z@L)Sb,y!'Cj;U!x!!rx-636k4MYW*_i3eDc^b4x~p& Wϐ+ŭ[Ǚ~Qs;p6_TmN4R%C&Қe$B*\4J)RmEy@:l^'*OI; s?۾qFS4iy[1FR),!>%0D+m<$\@v't=0&{A~ nn)!A C233>jJM$Z j[AZ'F|tmey{0SuvUV55)&cn-%h^iۛ2j4"{xa>CbG8|s5Qhh.q.2HHϞM fYf8d@ߨ!)v(P\oX7pW3O7K:]]$EG@Ͻyx`ߋ$NVn'S'ȋ۠$[K.LAaϝ;x yR~g8ݪ'NH"p6#V9Iq%ffOn"D3x `׸1bc٪6i*Qۙ{b˗^q2LRv6z,40$(4G&1ACxDY5jlvǎx,\5ml^u3g`o"zPG7(N?*q E٩Ri!+a6I̙ $-C( 7%e_M0bf mdGf9S(kܤ 6gw0eTsXfhˎN=6]1ț{0<*^Ea\qAַo/qpkZΝ;yk 4)7ո+17(h_qZ}<'.D~3;*!(4$Ը8ߞ=ՂQ@8=A8M0' B:<W0v&twQlB 4%.Y$EdnlMKڜ'1@i ZPyMCH$ 8dmPF2LF71TjL3Wr80hvEo`z,qDȫ+Tf:hS")(>zoj.Q\2 y~ Ͳ 5vn"k-+ե "m N !mlm~ةS]X-5cUk&>U7iWG ""@6F%a/KѥLi Tm3x)* rEI ͲV5ώz|xa!b,.M\@5' VYNƸ[39ldvPK</f-η/ IVx8Бj "+|J`\Y>y" n٪F2(ɁtmF fg/ӶNv$X2?mn~4i3LdJ_tLܨDMt:  LqZm8*QLգ$x=9qoo} .U*2gNDI79*^7=(ZWw=zrEh1nF_yZU&s8N~ϙ.ӡߑWn詽xQ^ W]8h#rݛ#/d}ƽ0ۋo8P3Àw/dIgZ<\B.8,.X*g,`(P&K]L*qGʖ-k GO:=W:4q3!l"@h)'SY0UM+ _3-$ חH}ϻ{E\:猾 pyFcy6c֠L $k㽑kT:XŸ֔ իe3' Wxn^1(e$Q|Xk+W7wE T>Oʚt ,QS/(5D>5nRXng;n3qLNm GK<r6⹃U8$\ qBb]vnAŪ?& `bJ J5W%JT(n'(C!8U9~q' ȯ!UYQ /ptlFV;׺hM $<;UuS9F-(,޿mFKQT)X7+͞=є/*Xk!4@lŴ;6JQ OXTw!"qnHʫA*spc3 T:7V F8xVS2I :g ƴtX #o^<B0H922\xRj rn%7yV{y޻;iw3ԩ^Y3^74y- MT(!rHrnUMa y,$30-!!!.>zە$zy `iHJ; @&WP64hph sS& A9mA޹lDT c4([9lEV;68!A2 _|~]jvL:o<47ʔ)#4xigxj :w-=I^z~~[nw۶ E|8\ ,XI.Rh1cbӷu!kim|+)3>ݓhn Zq, mI1vVHʔ`ǚ5dI|> 6u؈ XAsOSAyDޱ\@5"Q@缻&f uV\Պ`0(rߑ춣Fel۾Q>[8bnj%VㄼԨQU}@$HB˗dGqI_=d_m)Wn]4%4#L Gng 1CۓrΟ_,%MD1SmnmsB/2ūj̎AƱ=%֍ #E" iQ/A6=~:СԿF i*jnro C%HSQزisLANE*𧞝'] 8Ո # eTc"D7316?Trb!Zl`RbxxL֬Znf5X{ 2%DSlFA8AQLL^|՗vrý}ݣ29n`Be:Xv1r/ R<._mݡ`@KcK%KrN|YA4zf҆mӐrbҥS.[V-[2ژ2tΔk۸;fvJ>7HqI:98TtLcmBLJM}dfNC Qϯ&Ne)fe1.kHi*bٚG{Kax}س^AZ/vM*T{",ivo;d[q|mv=^#[&fT>{&5VԷlƧA>oƑj0u!5qcGD'd9sb09\% _u`ơ:|\jn^ ?}z ^/o{Qׄ &\kBdHי@JQ J۰c_W`F`R1ĸPJNv:BH^sff(ކ,cL 7@nbG7=e=Ydډw??qߺ284<(110!P:88'նm2nl6Bg}<{e2gHʔe7ov\~6ux9(B-d\ ;I4(;AMGA̙u,"%Osx`:_(>g= I;sԶYҽ-4F xB5lyx3h8`Гz,]gKcxcjG908kc͜8I|{"fƍ4@<6jV(vQ񬴪d")}%'[+fr\0ѭi40N  tPA۷-:4Uban)3 _a)Cj!Dq5K7kNGW ihMW72p^DIȹdzdZ E! ׈FӦM)aE._X5`P/ҥL9)>V]HGXpHE/.TMΨׯ[(1'F%K~E7ǁV^^۴I1A*2vLS@UŲByyyTJBdL :R@.aLa qZO޻Ä O0d vfjK[_c))&MpKhd2bceǏGnw<,J 1:"r3#Tog~B'#Iс4Hs2bLkU!p)3)3 wF~ႡqѣfŴN^ Xz5rux ={XM!mbǨfѬoɒѿ0]U)k`&\mԉo\&VWz7*K=߷(1!&6I%4g(HY'^PgT+` #Wl=|8+#I [(qQ{O U%^|2W2%|>t"p;Ƙrex3yscnLg[.<@CH#@A^=Ay$qr=~hѴ(R퍵۞-Z~ηn,QРtkk\ vtftQ}}&MPv6qWxGh$rNTQ7o+6*$Z8ANʱ٥k@[Zi@5ΐt&DVyȚeujir0 h4fJY0PQAA$$g9nNoaEyu{?ќv7*YTۡi9g7ΐ65[r^ɷcBvןQBԫ@ 7no/T͸ wɝBxXPʕ?;~_x!HC %KlF 5\sN%F%LEQܰ (WC@$G#81e IAH1AN"'^KP]z ņ9Et3k7~WjUoCGH>f7t邪W[%\Ѭ'2e]>0}]k?쾫>$*Ulܩn4J,h[D6OGL8  vlF uS F\#H'H(!h0PB<?`KF̠CaDx`̘1CHIa1E"ыzܹSOɦb+O|0PM#1*"x iNg&`(Wc_.ǿ(Ye  o!~x_"*M3!jhIU|iӥM$]g3}c#[Os/^6\l/E k&= 7>]W]+Wa'AxG:8C>3;܌p(jH\[ @=dڵk&L-2[1. DJiC@z5p:prJYnRkj2yլo$m7mйcUYpu/y槴Ç(ƽM#ǎ*[vܒ%SR76ƅ`^ [=2'VRGa@COs,֩S B˫@͵*0lԨ xEZQ @˗IStK*y) nbQto|r%dm*L$lXu.wzT&`/$X{jHLɗX+Ud(k(W~阴檔纔,gfOb 2b`?O(&" v-Kn ?dUZԯ )f!WS²|haoNL_`˖,$iNtl[>(5j&fF{ Iʣv-z )n4iXa_6hNu4$- r6i$ '*ffE.˫v4 .SlfMF Iq87ot͚P >4,Ww("  R6 .>hO4.ئ؎jV1G2|`l)"Ҕ>ҡx)lT7tz; (I.88IA2 hp?|*iwS')g˖-T/g1P1w>˕kВt̘g ȒiSa|vl#kc$'%@Gu1 A&oDsW}sA>CV@#4QN6J~` Fbmn@AGLǽP ?.TS kai 6IZDHR賭UMWӳ)u;4%f@6 'j,k%]K/Bƽ=O>I`ҫBOl5<)f%9^4bJ^ 8XbUG53c߾6-idC ?iRXU֬!CCFCswwX@sȗxHNwD :3Lǜ9Oy6m <wJ 3xrC8 8~',-$#6^m5 ԃz e Ýrҁ_~a 2` L}Hq$Ǎ!_? KR`Pn,ӑ#v=a$%1Xнd! A@xASn/_[/4[&q8c7 ڛ'= Ih)܅)Z1#H:;93p8x]G֭[Y܆s "1ԫgx1 Ik ޖEKQR2oFlho~y5k|[\Imnw` )UMpƹOO Oiә:d^+VI&`fQI*]چu"@0SW$jp8cL=dnLoX<(5{ V@@X3ux쭷Oa3zŸq׮'4%`jpjEZWTp쓺x+"jx *Sg9F< \܂HLe#o[ou7 xۨA\r,DJKTi{5-7Wʞ=;Rp(=(PV OFl}p$Ph3vX1wΜ!s:M/ ?>~R}^d?I(kQ8+:2 Μ1ln7ũSq6dDjYa棌 H_,yD3 =Z cmllI5AtzE_qZХx] ׎8˩0PI`C\DۜmWX@)ʕ tuSt }ž}zEs߽{$:.IP7I$QPz{Lp3ewu ȑ_:njM4|]~ڴoVwްaCP%ȲA:a^bʱ˻vᄑtFNġ|9<8T^=]ճO^K<>h7p9jE#H '>]ذ6ʖbN + {˕g F\=LπvH%S'#};e-c>6G[GtaN7ܞh'>Zn-\5vl0 G,͕(%k=bisנƒ $-$yML`Xrr@rSXG44N@a;YPET.:Ƒ@)3iYX[0`ڎz?R++S,>)ɍ[3=D۷gTX{(inтL)(tQ?W\Ѡ*ǞD/寡)g"OBx QO!?Wtl!n.jјsl$K"gm oJB( [ W M1p>md9MCŊ*guHCr0 y8ŊxyROqj 6j @BgW ұ+EMlhH~@q#H Bre2!WIı8$NH->RFg\8@\CN ?pnzlx`r;JN REz`J.y$ Ss)tnLa6# Ff?JcbYj9 "T=gn #ѩ8"j#%, QtTu%曯is,QÝ;ƍ׭Bs?SbKgoO?9`q$&1*uaW_)+oVs d#64+8)\{!J2׮]f͗_LVw!0qb "5mD搵AE܌>e6;b|}.L3Hʹu b##i3C @tu0SparxpadL`'=Fg0D:cwSLi. Ggˏ0&|,lE5F<囕Xq2L?A⪈>jm/^3q\PAuD H#_7/eKP&(m*D|$(7xf۰B\IC:Toolm-f>i!рٕ+FXb4 y^* իW:R%&MEiW_ExE\.WH62KЗr:4ZoCI0!=|@c `&Hzn³T>=pժx(>摾}ׁT +VSXK6EIVuiTv5X:u6&13K,ptS:ôiȚғX-HnJ6knL2d=L56Tp z`|90\j]ON)Q1J9:I4HqƈAbl/^$-YHYM ",@)Ϯz]@PATU?܏_<w VuqLͻvέ-Z֯oAUd,5/YD7@T۰ukszj`Oԫ~sbCtd=ōe80LcT蹞txm@QOlG-@1#IPN:j=S0# d}unfG<*V"PiHM>H!|TB/zu #C7t`dar 3h:;4uaӢr7 ߰[]6aE<!n09F:p,O d/ɇtGrkBeM'RC_ qs:~sHIITTC(vQ H$+'8C @bi,M!S :A-[:+#abS[}Bhģ>. #rcDŽ&!;&,857OwHM .oߎ,f>4N٣u_Q!#+CJl:^2{QHUƂhڻ ɰҼ43敊ĺre.`(zB[|6,gDqi-c!~7 ?5@3RE}A+ 0 0p@s|o8v|_JQgݺy*.,F5h"bŊU@+Dԇ*X7D/O>)BΘQCyyɒxzZlb`t>ǣ{H MұQP> NzNu_}5w0_JMpD`Kb9V44OGTs[+4__kE񈕦ǰ/`Db (1aeJ`LM!=zXc3g9Lm;:,d"GV`$rh@ K@ŋl!odvH͗tSER1 }.AZ@%i :QB7.rF;b'V+7HF +@qdY,T:p=#'^ -+k1`xe;]8.0Xi#)WKZ ɵHf$ Nw/_Z J/SK8o?^Ҳ7/WuN_}Fb6jdѢׯG0 (VIOd(UeqA}?\lt`Iӏ%e~I+|T![_d[j0?ɪ" 1߮EfHqfBsZ\1v;kJ_|mƌ@C|@^iR;rsQb@A&8O7{wRJVtC'C.TL='@1FMOx:iUe C K ;#;Tx% Z iia6T#]cRA%U)|!r-y6F}I &T(ʰK3>ր:&IeŴs121 (-@dِ;w]D=(<;\K>ADuCeBd mLXBFXZӞM[JBҥa+6h@gYp炻 G<9{ԯ Sp9B 4d0&*CF+XӡwL.go+ķVV! … LGpCYbk lr1>_>+Yd,:exK]v KML~bk6L%e.9Q@V(:{QFq]+sqs:+*'%tߒX{'LHJԟ 9gaEwR 0r(_.{Pguó0UM]/J=Dd=3$6lOTDg\x Hâ5@(wÂ%=Q$Ns?iHzf90l!sE|(Y$$%"K/T Y}oGߝc mssE,sj0dCҸexw:q@?LOa)wCdMD\AXYm |' KرA{b =)6D8ĉh8c'0ƻ{&.cpőx:o +U8plŤΩe Gm58b!D`i\ [ou\ # AHs`^I?glPfЉW K-[gYLM>|50׷p#5:4_iqQՔgW~0|9'^o`Bsyl=W^ 2v2[xvϟ+F{SHf7i0K+c ?GnF-l e]3*RڌE<J\G)66McGN #`\A,ٳ;t8Q@:l,:߿0kW7˛)#j;6"NRq %2#4*I7' : &Oө-+!Qbv+VLԛe`*%O6=ij3m&OKuWeى{Q#kdUJ)%`yުap?fwgkZtdsڲe0wn5QMOb+0q}3*Ѷ{f-6b >0! ),'#J%6~Y:>ҰmR>tSjC8ת@ǃXyy7 Nb3ȩ&Cf@6#q~$qs5p 3SCW۴ qZa qH1?tJH3YcK1l#A/RDӎ|*#lx- jZY JMWRKc SgޚWb{i܋r.L3J7N;ļ?v8/.tpj.-td yaa|Qͣ} >p hvqTt/,*55o͇5Vŕ- C}2(5cĂ_[?UHw;5zq[ܸӱLѣ5@`niUxX!$..l.=WmYr.7:_fu#tVou_6rD$ҥ7]/6t<byXWۛ5_Z$`<)ΙO "+O?7h';  #x&G6WՕ?-(%#{$#L(viwߩ_>wwvfQk@gYCSZ#?5%RpE媔 ԓFX BX 2-/QI{fBm4erujK""Ŭ:+51kV{VxxPKx?%TI) UR# >dM:U=2$p"_TǠݱͅIH2i(⥙n'ukBNA#ܒn/&߾y&ﬦ&8݀/ փ1 x`ĠM~Y!}h';7AXQıo!+t82PHCc%T:Tdd,Y ɓ׮2G1b;.]²f/j22Ei5̪ bࡡmǕMg6ܧ7$݃-"Z[H"7w}&Oovg~sS3Vl<6=|4fr J0 o&vg~$ȥ>=:̝ٳMf6:V*VQ۲%N$cLEu-\]`oEs]t&>_hqȱ@ [E x@ ?y`TYʹݿ2zʡكFE#=)=/Q~pszW?؎ ,/|D}޼o ycG,l-Q5*~sBQ;+3:IiBt*Iyf\<,O6Hҥkン}ҥ,.|KPiQw-&0 de9wC[(Tf~N?B )14#ӅIpPp<En`>IDAT /DaAG+)ИLO+@}d.)Ȇaj8#p8a)޻+w>Xx0}!>x^Fc^6vdk̜dd5Clϟ~B2\@-T2(|p s{F`0#zd6:^]v-n/vk!)Sfcq!\@:AwA ^z)=֑g ˖- oO(ت1čE3uc7չsoթYLf@ܹCCBJKB5K ȓ/T+4^.~@FUE(#0xcmzokj .` 5t Y) `bŹ1RzY=ΘGFKOOND6oPab1unvHݡ #ȲsH˹\cTinLl0{ ZsMMN{ZvO(QHZ%AAM%3.,gj'^ P$HUAV7VdX'YXw+\8z,3^Bv|x QUo|`ٳ|S}_ϕKvQ!gȰe/XPMk&JZMfiTG^~=1)XBR-.%( _oPә_nޜܿФIU6lolu0A^^ >4pgªDN7) 11hP&1@ 8~pO\%xur0.lBxAgW1*k!RH'A3Ys&-Iךޠ&|r8klȴ%aeRҭB2g=Ν ޤׇXpw·MM}aavO?ڪ9_>%iJ -"!5q"6g%4 )'7!'/,JƼF1@\Zx܎&ţ.y wgn3m}OoIU2,c&.gum D1{,}NAYA֘?Լ&oL>K?A3h ٜ*ut̛7e 6LLN10XZ^tZY:JRN{U(c_gkP8x`u$S~ɠ) V7bKHH$ًFH{čÛ@!ۃemq2O2%JY_j!deJÆH?B +(YCĊeNvdqF Î""lA2 O:N׆Ye۞=~\GpyX ʗԯo 8~e%G'RGlzf00C^&U$8^4ͻxb0-^w@ZvMnJ$aVaQ/:gn`8qhcKmQ)k۶%(@EkHծ7kUFZ+`G`wD)Vd) )ydOD#R%+:ڴ ?|DP0,HC aSO!ҤL̎=;2_Nd!;R%Hɇ'p s IWqfyqsK'O`;|tF҆On]zʸٳ#dG[1>1y 4HV ĝH#a1 { :_Y%@W6Ie Rl?-?Vj2i55 f<}:~bWV5LL$`Q-x;)Fe2/XTF}nK<$< u ҇q#D@X3uu0t!6l;\m8ֿ+]#rGϨ&  Nأ<"+/M4ԧ$z!jrk;w;wr>\)X,@xy׮a۷+fVM@!9uV7gQ1|E??/$*&Pڼ\X+5Yqw!%-;F V:zp鐈}яf.M8h e,)v/i-y;tnWtY\r{M 3"[ D[.^'7}cD$ӥۣ/j3ہyB~-!O߾e!?=LLC\D:N5ۂ؞Rkx;s!=-LFI L:uǎM7"9嫯=A3<*@\Vo6@yNB7rFJ*ljB5\CSh6W V*B\ER |*ջzՁ4L F?Nܾ+)1X<}X\pLed=@˽ڴf6 G3R,1_5Lkz/,(b+g/mglBKEż}lϔIm+jf[Vo+[5.#{Z08E]÷4g8!e Jyl͇͟G$}04ϔ}l5:_jW=a%n Ɵ /"pׂ^pp)p'{IdR } ?$H-E78fXMl 6-^gD=\$/ z#mÎ"7ϗloWW}Z:Ta:!M.1ia$Sf0]T_QI4_"@$7nrdrlG &h{rqRcfΝ;$ijժn%M -󤲕.]tK> 8|r}`%qȝ6.g"vj4`A? c$T68 `U`EׇH tGH3bH';JpJefI5܎Įi1em%\0<֌ 8;qQHzFTuKU<HQ]Ά Ou4<`K4 }:Zk>6 #*FCTx {F6n NlE*6EW=t^TxC$aWDYzo]M}ԑk3aB#6lq 'iӂF6jTyw%z"/ۏ`Lڵp=v,BsӦqk,_NIIʬ5&%Bav݁~) U0έe!x"ao%|xXҸиYpR1qH˥ Os1)rI&,co}lB\86W :>6Tλh ѣGy)U8Ml\-}Sp7eFa䳶EqθfҦ3g`; c_{9qvڽ Y\wuG+&!:T=qMv QLaCm H"5q"ܯӋi8b,M_9^СdZ#5kJժӧc M6*w~ &W:{ lÇ-V7{cqXIi۷#Vv^C9N_lK.N]D֯V°A8}#9nw&9y796K'y2QL9j׎(XPNgɒQ5jDECa"(:{lux+pJI$-Rx;a){UW#̌X*.TcU{rC!bA]#Cp)v!ea?}J.6:SyCc&8w:e=~E a*dTl4M՛oX6*gZΝ#rE\}ڿrӧOܰ8.M<{x8/U /QQW -ҭ['K+V+m5B*XQܽzfJPHW_]1DHAҺ!n,rKk{"r\t%)kKީ-ݿ<+`)=gٲ-ʊ)bSz_V6SI7dAr qr)))Ap%ow)^֖>N꩜R8G씙JU E4m} 5Y-5kR/@eNKu4iW@m@t~Jɐd&A5R=2ԷO j`'!vF] TJTa㗆܄Ȭf7S{2y ( h7r$?rո2eBiu\ld9*ӕC5"#ar]B-l 3_ lzQ"x pFx`94`pB:!w4IՏ^y%"}.(HP8Wy#ٮ]T,i/)PutiRra 3wߖ\pl30}^z)+J@={k+NH %,2?PMSJnFY&$, 8ypCQ {ZT8>4{w8y{m?40l`*r$V@e Va-E%2O+L[d:L%.-KzE@&Cy- [%/EЅw|ヤ`gjРd1!I]LíMnIlSdu]58oGY-B̬0o z F)4q]2ƹ,tdj|t<!zQMxDc<S͛7Ժu4o>{.r>dײu[`imSO{`R;*RlU^,'/[+6N?7"o!lٳg\At҈dp H={)Fz8ocԅjHڣ8\bKBrp'aN8YSN%gA ǣD JSEZI6[ٽ#m?ȣk*"00819(n3ZuWe M g拏;}qÒf8Ow'?IbY6ɛ5?زPz.FEΖ.o-Z2ĭZ%AS+]|SJ YQfBc.k6!VǗC0CA#盼 NB%Q2"NXFxe'i# ~?(ULaPBB ma&] n|gΜ,Ÿ{#15cltv%%SOY>IT)ipZ#Vyb!9G#ą;F0XXdc$@k_}X. Ó04EF}X<`ߦwʮw{IKJYKoH97IKD)$J S9 汱;we7/ Σ@H8Yrd c3.Q;!f|pL m%<`qEMw_/f60Ա}U1egw dҬY/}S}/ÀTcqWdm^>ʍ9$]gWGъ'= /<:j7o~*՘'D.G$J$p)4*QW񑤀HQ̹n}qRzrЗ!"ib0IZb,p)E˕6Ӏ%[{,fgMK.hspt?]/"xn4ƥ"a#NpXә4O5$9`˕DP9F\!W91tdP٧=./'p+?sf޽]MYYx<ƺ\{MøERYҋ/J?2061O/^2Gq:!$8Ʊ]Gj5GҸT3V槛i'Of$ެS?YШ}J$jjAP9AA!?˅*Rԛ7'/mhQ)~*'">Xa0^50ReOҪҕ6t({ m<b2ț!J}#C~/QT{9LSɚrHn1-[䱑Ew4Kdz 81cay$EeG|dS;3D&Fpv`YcJmܛxkK$+:y6!ngQDUv]4IcGSK?TM&}-'S1:uy!iQd}ꗳIƞ=SzmVṢo_yY07SLI#$]:CX3vK]2-c&ɜbg| it^|1 ~5aB}ԩ({`孹+U~uz= CJ=H@߄^,u0?@`F ۡC~dD= J=fK/='y襤h3N0Oܒgm6L3{7b^5hb2 MͩSƍǍ&1NJd_”)\d" /$]TRT רaѨNj!;&O|qdKשc\YfzPjׇWٲe*V#H&LKq&359m·n]s8c^6"8XB3^)D:ciOPk+coVL1&xLaEE5]47wիE1FxC ia-+GBUw }ջw!Up\nn:pYƕ&PmB;6Ƴ(sZJؼsRܹ X+M*qHZ՚2{v 2/pKbPVHȾg͖45u7@JΜcrN~̪ZҎc`u%e͠}EmzǏOZ KQcoS5+Maf{|+/d7G@# Nr6gH\D\%Ν'N)X0af(g?w6+ϙ#GD߲)E- 3wsљsb,6u͉Ƅ6k $sb}uLK*.Dxj>|1Yp\1_=ck>XTUӈ+diOo)vEW6|V^$2W6}rJ4`V4.J{{up޽ԐF#01%ec* Itt2^x ˛nY0i~XjWW[?x.#F"3 ?pgx&T^T=(uׂJ4rdoNK[,d>PZΖ+۴rVtJ1m)6~^nẉF?iS+Jmd_*[HkئVIGR)(L1)pV˖ث|n=$ _0ֻ n ;d4KdI2 V1=ZiJ 1HAhi7  K@ RhY2,^@.#lHO۷RZ!b(͠S5 Sc z=Uq!^“ KȜjR4vNѾjUⲔsb﹖Z@ pc|# Ug8$)͆Li09`t/FRvcBZ_̌=@!bVv-Mau  7]v^[~? 7A䠔4Çr|JdmqZUuQ@!QTȸ Q2Ɯ[}+W, -ynŕ:r/_b Fcꖬ|||p^;qGepo,2_ʕ>wxJm)YŊY>o̟ҿAg!sRۅE:)SĤjfΞqdտ%9t^}biE e$ ǎdL&̹<BoL 'WЂFS X$YjLv`ΑR Y2((S'Y8?(EEdE4"P""!`'}uTuxKX#~B=]SFsHh<~;!?E7[sfTI"c vNf3# )$Z'Kgy`/PqfJR}AD RUXiXXʝcca 3'#SDDqDR[Ď4`RmSB,|-}$)A XZ}ݶ[uMO7H %_kK3(YqElseA 4 -Bɭzd9!kl@ jwz;gbFƥ*7-)<$;_A?~~[-5H&T'j3Z}jVMrS[bbu / I ((֩VjVu;ڌ,#"3B,%3M1%S/H52c;f0;'BJXݻ٣Slz&NTœ=:gg DxEbE=ģ\7i `'zjvlBiCɗ/B3st~0L4"Hxl.Nj#aP"AD5U1GTsa$Mj˖=DtE[ȺxA}It$[%đ׮d0¤F߲e6]lK`3GU T9ei5mӧe_GL.ё"aBd% :X62x 24I&Nc#޴͚\ʚl -])..|0gYEs䰱mck\ѩS=zAhȕ/G0+ -ƀ;Ay<W1@aeWf\LME,+>hȄ|Vp$ht`WHv]D"?LNw,v|{Nʱs;B#Sn5kL>IYCK/;EO+^=gt%hHo_]8 v014rd޽~Dm˔sz-)bg9hi^N]'Z KGx\Ξ۬ Ga;r]81܉pAa 3,<=LO?|$Օ-k?3\.ST?kxjՉk, xw rX-= p̿@+;TE]lv(Q#ĥ@tRϞ=5>ܴf<:C1)6kQS.\ =ME9]Krpm5j=w)NӧϘ1 z&Iس ȝj9kV{_Дih.'JW*?k r)5g׫ (M+xϿx1F9w;UWX:'KIWp2!)ܵkW®d؈ec$GɎ?ojޱp,>K__-~3foJw M.xlrb2)A s#{wWJZ>Р5L!P^ ]xٙW4Dq߽#4y {Wym\I bE:D#EEvRnI`FZgTEԥ׸( Ñ߸1NSMyjաīCQ꜄l-$0t>8|H]FD9럮5qlQ+!-9O,OR}z0MN# EaAtd<ɪD`!LFlE:&v`#R*U\!AA-k֬v|m$ o[ ߧN-Pe[}imC$X&MԱ̚~)ϗE8軟4R˟ż\VV̯{|lM:HYȠ%,'R'|6\8X'6jV+Beć֬zAzDj7mg?QTiUr 9Bょ?y8I2j?_昵!/8=ޑJ'HH_cC &w~ =% x7)EٷbǎEǏ 9z͎,nX/' v邴obWuDjj%o*{،+vm9T)1Jidal0o$#U$hm[[<ߖtU2jja׬m.I0_(<{)3&ٚ10%~"8ztlL@9s9,ә_Zxx(5xlMՕrҶ,&h֊+ rB'* Ѓ48%]QbZ,.`qY–鲸Ua,:jbғѣGP&q<*O2ΞEDNIzkعs'15ɹfz8-wE쑕ȥ'+PLi[Cp}6*ϥ=EO$oժEv2k@xgM$]@겿K?(²–aݐW?{'x|}V.V24:۸O 2 P 4 hhdݠMmmm":q F>0A/KIρdV HM` nR7~4$lTp)w!ZXHs0cܓq">R+ess䄍ZB K.~(~BCH%hTĴKm[(IE*<q!$Uc6m@#']mmDoʟbO\ 8?W+wTsU1 |2 O4VfYቊ"bT+c!:g΢`v+5pK[ b΂Nhr|#F0a@H)rzU,Z5Q=+{plScxI#7-kabosV-G1s&M5~3Rj텞b 6lU1ĺlg 9i̍7ҳ]$'hA6vC, f=؆fLi 2=孷}4 %`  &,FInHx+3JY¥<պҶg#Oլw[w 3dQ`dpA{t?F?N:#E7Կ=W:r\GV8{vQ?_m5.񦡑NPHlΜ6·B n,(aMTqDVZ }ė[dNY9S 2tVݲ}иeDPjhvLۜWn|=2ysc,a_?9MňRdUйs&#-` n Z5N !r.d@?_sgNr*RowǞӮqV$ 9~ a."%yڛTf$7 ɜَG*ޟmx("ڟ}w,[r#Hw4Dc/1eKT'ƉtO?!^_4xt{t}P*did{`c ,ܘ{ SIdm;pֹ=B4F\ 2-f:D;۴[t⃰\P]6-^h-"Sܹq&j,QY+K_&04rxq69i!NGٶm8G"gA:5ojsd)2v T7و]"I VY|Y"6<|Hװy[xj3N[]F>29$e'Tx[2rD,'H +WʢY>C\7o)alp":C^j^*‚LPI؇ 4t_#A5@P%!07DLG^k|mV|w;>NK^@BY裏y d!v :y'TܐǧM0l¤AhEBZg=*TctaDGӻ{}gL`߫ÁċaPWe+KÎi*)VDikGfϮoG Dj@TVw,Wr `nSd;sy8 #]*̤\#.[Ʃg KB^F{%j'pf۳ވ[/i/:o g&?;k.>;~7! .PA8K*Tr|U<9v|XGaBXzB;ZW<Ț6Tlw%wJZQf3$ +nz+A2*s"ile0aGŹ융쐔g^\aŽ +677'0#s (LScC~jiRQc%XkOЯN@/BTfKm"E`QNs:&E&MWGBn}pm'.]b]Ya1O0*߁ Q)pśI)o_6boڬ;m x tWo 'f??AD2 @$>}dz k!,aBBX'՛7rJ^' ͕34>qV?'X1t*Sbd {B|%vq?> 6)Ufo\ڂ342 -oJJ54G*&-.i٪k| Kh)C-1C 'K*5VpJT{=TbYQl/$DTj=sOs10$쁔Ih\9rB ~5c5k&yl(%8Go.qn^Pi}q\ӳ v~68W_^Bj&oEş͍p"tfM9>+d>{6ݶm cbNlcIa.x !\68@ΖuTZqY;9m^Z #N >3l5RV~qTqF-=|v1Q>PZܕmBqpě7Iu8m\D̵9B-uk4xx@j&ֺ7(+) Ϭ( g3Tj!NJ$$ s9%$=|ᚭ[1dR Hez/(Ԣ%o}jrs4_Jd vkot.<89ٌUZ, ǎas3C0aQ6.rЙUvO>f^h]!@\H+|Q\.88l~<رU?E\Mᆈm0<^MG#I4a8- q]$@JMSB!A2a۫kBȎG~e[fVSQW<aohJMAoo9[s& ȿ^8=is]ޓO>)7#'ܧ~A84 YkstY}Fmόה­Ez1_ VY,'‹I^@v&Kϝ,6?gw96j}jTШxmpڿjB1&o'O~Ia̲%Զ0쒓e+K(Ym(`tub  #vW'4TFXR۶eZm/KGYJAHľ爡ϢO`±0`+ܱR5M.`E A1'B N W5`+ r`mtzI+e}Yj׶?|XDU愺HBwIxlkX[=#fB?ܯXs^o7q]Hx tܹ/Q vM8qb M#MJWM_O*%*b kX]B 3Q'5J~mb*| ?{2>.?~j.ͦ9t{+or^'!{öX K=w\c \ {]{%g|XAGQ1m5y@}V<djL8paBVGиqGl4%;w Oن|\Sds`ܸ H%'J>1O7L.[l[}SϠBG,An''jt,Μ1m;؆6PXblpΤZ?'3z`5J=kWˆÛ&viA#Gw$δkqK {P +$ C>2ٲ~![6*R|q^ 6lvm^!Cvlj6.6i"ؔ`5!5ln Vxa._60޻`3nl{BV 9BځXx?R>$ 9"[M]|'kҮ>-z5O4q|+qҤhK۠+dzp1I gyC:#P"B>-:KF?y[8Ӝ@@_QIdB "|b?~va{U>5߃! LWtSML/ig[h%g2|(D|P9BНUp<>jD091ZApW:EF=lPrqDng/|},(Oh{x'O&Sv ԩ,[놟2Gu9kg[jk(sE'HrzoMh*">_dq,W͐۱~6ֳl*m;%c3_[=lgs&i ){ϞH˔^fŨIhTqgwԻ]X%JD7P:0Vmr{@J}XH_O?mAFgV扤m{|0g3EXĞ R!9LL_!9h i]ZߪjM z[uM)TTc!Bn2endd%ڡ{+IluJc^q W1_%-OhӦ(pR@[r "gճilǔy-q_M #ڤt1*/as$ʘ1 FxW ׸pu:pT,ӋzRJ;IWEͥ,4Ȥi'мy0HlÇOL(hQJd,1عH^ Lb퐂Ȉ¡ݢę?Glmgf|xreu 3Ի1"Vx8$7TTvOa#:o_ˊg\4J["SO9KN냷<"lϳYo,vV,L{t*흥{nbt1O wma E0S1;ߚ׾E@J]P6&uRuUOntmTT((9Ɔ8Δ+'*l!l ]J t8dg'B@xyٽ6I5r(G{7#0pȐ ,ź ԑv36/XXq9M(.f;jT0(X g<#@9q4kIdLz%׮:+&t?<ߛ_Y!l6V!C?8h*\EW46E|i!7n{F-l68s!VF7絸9+Mpw՛{ Tr0d^-K9M7jo_>dPT|}ɸK3_?QL:eg fu#k] UYKb2ߍy:U-ز l0jwKU3LD.!f@^΍ ֵk(4OWy)K0I{J>,H2834zF썚[0H)%=IQC7#{/I(֍`nvm^@̇U$jR]/럻3Jx[*e\!cNkq"Ŷfy"7'?ZTc3)>XZM2Yw y5X!`V}4… F< Q:6Ђ!VcFVˍT0tVF~b)bMY?2)xVl;b2G L O'fޮ\c-,[IRJ5%*1g=ƒxE.>[UJWjf>X}K,]ѳpڥVr0,$2|x0/@TT zg Q=kdBc?`ŋ g у̳8.-f@s]Y=WtwCva~,PN :f ;lƶ hlzH ] ü!$^\$U3I>+xr D OU zy+[̵Z:0z.]`Ma`yاO_KzoC08J!em5` !kdؼ0Qawt,.ׅ-sO][X^9 G𛢿Gq6쿚 7 uQ:4Ou*ǪY>JD:O 5}sMĀ~҄`  0{I0eڙ~fjJ.]0|LfPɖh7I)0G6kSۚr6,qHBnc$ |z^[D@4ilݺQH$R ww(y5M' PwHSjk.ҬOMnlߜkBlcÆ8e#&Eţ痸wtrg>&Z,Ad[ܿ?K7Ͼ8pw lRͪ-GE!1px}{Wǃ'E,W`o*X],!rJ>}zll;G?Ul뷷P<FnMx ~;@k6O:IϋI sDAyrƷF1 `+h, P]w&  I8,j( 3uFP)'Cz@KMΝ3ZW5n˖CP04uqQB TkQKȔEB"ol!-;_?uxt8ʡ5k"{Æ%Wji o@G^-k=4"&1Ih Vȝ>M“?'nb<"+ҝҊ+gړ)8aTj<T[xNЏ뉝 µLLŋL}R,Hio:OM1a:Cα(O @X&IŏO^Z1"d륦]d-'c Q#֭[*ˆp1 ÇuԠf|/_{፧FtNq  j2d lWي뀁y3.t*wˬHlV%#‰CNQ :Xَ}T=V".1R=br^QʐNi,HϝVJ"qiZ*{ i/B56GD/>K ΋`}8(`?*z3q PΩ#ŌQ6=hN+!ʟA:,iQ3,.s{ܭ1~EHg'۬B3 h7"RV$@! m(d!źfvϞ=AVF EɰZET)*^$u`R`iC";T `BC/-|5c>%C7JaJ?bˬ&j*c8ڻ$;\!50p؉Hxثy{u˻E8,$ڼ(FJ-7F`TpKHmAĬACW/ 7=~ 2_ 4V@"+ڱX؝, %z$dÉ(Y҆3 G_w\!6lYڳÁ֭;rw16䊧<Ц\9&_h=RPyDٽ#tƍ{8qAAaTp)N}ڮ)f'*kBzN")wmR`??Ph\}y5hɷ^?]WS ` H@.鶸7> cK#{2OB#z{LJH; 6$VJ8eiЫ~L6б:"wP}$Ag$#Ou_j~+ 8fQ%"Zz_]a:2z>苋-Ɣ#-@w°8OE!l2/ oVX]Š0}?罱dV Dױu*gA>±7pV!m dOYq b9 ,٘xVb Y`M8Q 7#FWeXd Ib,-QI,S^:|HG1A `>^)P:No'yGՐD4[xB,ZuQRdl׮Ku REە+WÜșpC7M9_q. Q/|hע7wlR۷w^^)†V70bּEiVdEXl =O?0(E L3mrhް1 T%`ò |綾RZ>82Mҵx۴ c6]-Azl5MLJؼysop!RQևp$aY0!٬S$_ըQ+ͫJIC +rE>fTSAFeԒ6Q\(0r l@ѣG?"L("d|FuwIE:irwy~嚤.G;ky=;ee@֪ߚzg noZ9LYɈuƮ(ш(=D ItCNej\R!羘?6\M`ܹ(ׯI2!6n#`Xx~$(XBJ3X+5f$" ƶ,"@ \Βi@WX,F_L*.GɆj/k8,f+G0RHMA!'"ٜ?~ +9N&ëE#\)$e6dzlsXQ^~+g{Jb*27m<%&_EۮN3-_4$e_53c12%bW4E/81e˝8K&C$T8z jyB%'9{lxwp]FAFm4Xxq^V\Iae)Tҙ&jO@W=1k$JW `)r!N}ԥK e#d>pOsG cʛQf"Ecc|\JCVUZdpc h, ]ڏ$q/HQUyuµC N ێa)>@  ep~Љ@- Hc#O&IJ}&)r9X}PـڼyiTwQal{{5F ֭)Rfq# X1o2z5+t\W~I{|?K9Eg _r^_Zh8X8p!?ӪzhDS#Il'%)pnヲ`~zgmZ%:th9 L׮]2bgjYw،ؗ,у }ѴcY _rxJl4!@ R!:fOT8 ]*KHW3gQp cq[ɧ*!Nc2䀿' EWfDR]PzM3lP])RX8QqĖwːG#!*hc5a N4Rq+W/ )ߜ4uѲRâjՄ=7"źేaN09*y<x)3-R^}U_~lYqj)( k`ɒ=|d1_N_d_|e@F@8t6\ p(9"D'>iX}J^~uՌm1p,^ZB0hL4"Ɔ 3Y:b4|>w#F28DyL,HCDC9\ K"?ilAQaZs/äxP-Yn ZT603P1 0PpN3扠Yɔ=ͤu/糸tY! %ƨ~}39pm(CI؈XKY 1\ܻϟfI9`2Ea1>$aޢn#y CݘVk7Qg,=JYMk>$V|fiNG'~d~dP!M Qj4$JR֜R,.KT& $~ 8ٶ 83kz| 6nk@}XDF "h{e͕Z>r"uGz0vгf%9S{I ; g%7blgcOxc:.i,`Dpq Z|Ot`觐Lkޮ]Sϣal+7AFrJc*%RSve|xD3&f76Պ#@ƊDnA \5!gfU1MckT*P3ˏM BH iLQf3#@+r7 ( )Wq-RtZYO Cڱ;_%F7k$j}:Z֎9~Mۉ? ,dBCk2e Ax4_~6weC]=Dtx=J.*Iš!Q*P9\K'pٰ#f#>({Ѣ^ԇPl`ml$J1e H"kƖ0S$[ISO /4%#(ͳ͓'?`͑$*ΠcK*'K}/bƈC?HzHئM8W=Y,qA X J"ԸE1(Mbx򫯪2!$HRbiJʇ*baf8(wؗWǡȅL~™{FϞc*T5@012Cʕq:9e ]%$څ e0n1t&*i7۔n#D(Oă! mfg2\, e$uOxDLȇ?[h߆ ldmn^)a80a[1ts!zo{WDn~p:^Du>8(jWiٽ!A[GTwN`sMx#,6*hb"{L &{}pJ=-9 CǍcbcvEKԸAA&HǏ?|:uPD]קqd"Uo9j2m>?O;/Q3Ȗ$q 7#w Z#һGFdjfDT00J>(|,cPDV)4³TRO?*StBG꣦Wnsp lǰ{}U"5 x|@Q1\HN**7qQahHӨ0h))Ҭ-% |/×zK-1JC~ sV &\*d{5,9YQ2KK_gf5^%iAO` f&bX4 Rk/yC"&:^q!2cߣG_)ȃF1';=E@`eիWxQ,/IrNX@ }E4P)t W^ᾼ:j^݆olԆbKg*{ёuXW(2{M!qdN+S#TXl,4j`#B gxr⎧:Zլ ;cm0 M5f̟)!$񒸴cC]#8lkr$!x]gi]7QD7XcZ9P )eK p%=hRW%Nq45CނY_4GiO/u\K$[eh 27qWN!I-[(xC$m'ަ/_"ljF;A%;w1HYQoi&K pQ:|i~K0EgEQGyvT04DB%0)tr\-D${ڸQMnHx? QޮtNC7)DzE17ƍ;Qq:Qoq)>竉h޳]^^h͞C{0`pQ tP80Vja{@#甫8oX&+vYpBtd E(ǣ?PRunn~?%|ծR:n‹ 7?2lrH̓1bajn܄ )aBp ӂݭ[񂏎 b0eWCG;[!m*;vlK'8·}..`t@A#\|$(W*m@7B0:uɄa"GZ5=u҈Fi U@]Z%s9 L"Gǀ=]36)K%DHO35˖3G0',VtQV1ǿiIY7<,^\V1lXP1h 9DȤH)pfiaL}TIpzTB`[&QËQA( 4KJ5P=hnªŶK =HHʖh6Ա]=lʬ58!'o[2ץS٩e6ֱ gȥ/_sdM4ɖ-i~Q:E}E_t 7nqF%GM,rL]o8-Xå9vWf&x-B9ߤ_JDԬJ1!Cq2VbzDo_znlr̙` pl0 c''K-BM%XAT2Ř#nJr\v[b!,@ytʵSڿ\kj9TKq{!}scp~u/6h^ Mnh輬_cCQm8=QE:F_b@9US6sO*{6 zHT\{ZLc#hGQki~EF+C@ ?A8(8wYC>C9履d] .TC s{R\$K-=i M橘G0K)bf^ng4=Tk )IE09.^4 1qƌ\/AB"%MNm':)(A}R* P%`slK˕.#,JbY+|+#qݴ?0cKSAOn;Lˏ_ju)ņG`@,,N:/F,'|?ӡcћdb~o^mr6Bn㍈@@.@#`-TTߒ4Q}kӱCxAE;͛ P?f{e?κw0cNm2)SPZZFb8)hX+0Ab#GQ@PVY|dҷ_(mB5sCAݻch_Pjf(qҥF?`tJO\OfVi /FEDؠB3X1BH\uZ^xA{lٷ9Bk%`]b/, 58? -a`szd.B=ĆYo_JnAx)6yeúEeOqc" U[.]p#{u˘>AkpwXA/oS^>XRcH8Q!9FSA˗ ɀa=?TyN|)Y였چXKb`6~;ۊ._&k@7=g۷ odun =|}Dž G$3E.T7+㏝-h ʻid G2Od8cm5rܲe޽5s*$aRn>'C&n%` ٴi6%,gMNEE:޶;e/FX Iw*&_KV`M ^Kysp۝a:ksPssGj 촠̙3 jVC l٢)d\8^ zMj'=6=JN(j9{춚izs&# i͚X r[sݺ7Cl~oo'vL9dZb5{!aXC# ;p8 *ٶϥ^*7@Fb]}F$;JkPtL8Ԝ 9F ;˲Zi"#&ue+/`M`K?,%{LLpR"6m_8'(r[JSN@8:!J&K93%u7>vO=]hlVf?Cߢ}|ȩic1tMǮKJgn!,*acRb}ԢGg-o(%"{H*ղ4B^,&:C=3~yjs| 3 ո:DO;.J E)ؕD+^b(/x!UOuꄓ(R`#lذd3hP(is ZaB\xO AU==tv@}gi˜)IePpu )24>FDb^{44QA3~|{,jACU\e >Fu7H>\߼=UR;~7!s^/KyI7]2ȴtLl0E,g!6 ]7p"S3k̤16pwd/y3ǖř)ɘ2 tc1HؔDv}8Ks 6oIg y 9_z)]8|XA$8JLKjtk=ur6ScqvE]߹LEY| Uc?/"F>oz"ܷN"O!xS|$ 1{<˕/R5\I uCU* jVLW+HFX% CZ?07lv-Z2|UhhgyG{"}NrE(*i|Ж9!/ "%U!dkFqiFr#`6_|.vuG:w>֨a O0X?j?⢷%DdP`@H;ÓG OKސ%ռ}vgB'A#vL>ȟP`؛1Qqe6xrpdm|k?h; s/br{ԯ_G)dB\!`/Ӽ* pvbϛ B$\Jlrhū|Uj¹ߤlKPDqRӊ}=??(^a:Vlܸw hDoLʞ=R_M_UtWK+/UڙϢ;: $wg3ӛgVAn͚KÀ@TL! NWfc-Fۂ6$`N7FW0īPGOvx0^ DœiQi!_NvbDžnH jczYЁ$=hʔ)wO0u+WcF%X[߷RJCesng=#M } } 4'OKC;`b|@'l9m7Irٯ_rW'I YrI~1L+j`гgJd)NTtώ:#i7 @igz©SD [> ɓ~3fj=Yoʾ}OHbb.qsK|7ZD}vz0K0!>˅ RrE66QQmW`OxM?o g%Ѷ ipR HBkʢw롍ك qp)adn/iի[pBEECoƉv'| fҬKjHԉ9y$/ H#$uk2&Ŝ=azE5O\㞃oe+pECRk)Zp`!k,wj<Т'gk GHW|9lPr85=2tFH޺BKyAxxH#`;qDzZKZ:p3Ϥ iMwYY:Y i艐85jT F h&MABcSw&>= '>z"4X~swQy@>x,Ate^&!bPDDfiB!EW8h`ZF"RF*|p QMU` yt}:K/OBMS<.nVb>;\mY<'|gUW?C}0{\.]CB )HV$yX*  c T\ˆwS^sQ 2 9^sWdZq{`aHQ̨[|A>uK-2ۜ*7bAADf 2-5?B4rmln!h"xի?ץzڨa(! <ē FCmkk=(;ULo_ Z(eڵ`qw j"Zs}v;&/` ~NDIj~AR߾Ʌ G1%(5V{B\L,Y\{IEaݺ9p,$OQz-bd/%#ϒL.;Rr'Sܒד~.? ֗vS͘tܩKy#[Muƍť8!uW@4ѣRD/Rj7 $xO2e(h#Wz0E1"э0PmF*PUY?fŊiW[opLB6bNaa%+TpaA|(&Ӫ~ .u.ɡi >|ObT|QZ,]n$yilϱԁ]+qȿzU6yq0" m(f!8kVdt&ߪת$!c,WΖ'Oڵ 7aCJfOrmtjԏ%fU 0M"pQsӧmS n3\긵d,=DwHzJqRȫHiH%bo^iV'{űH~3R\I9`&O]O>@6QgLωy 3zsE͒&BDsmj-}R 'o(4!e $# #1PrE[_@(ha@pX?{wyQ+   '(!"hҷ~K)Xl;w$ѰaCT!ASXɹ%߰`A-olApӊM`vr݃</H;s=O*$j>v hiH|'ڼ9)Aj a> Y,ŕSZ[ϭⷝרW 9ֽ'":@KJ*yLqO7vig.nGFHCT?Jf> geMąըQ8k?UKG`Ł`V%C3Ag$)nɵپ=vgM֭g}m}Zxq2 ʖtRpy"X%X*C-]|td:5ԙh֬y}W;$>,Psdݻ|һ&;Sx"fܸQWrx)GIʔ/oNF}^=j!bBJo ;jc;0vl"W "X31QN$gd*܅JW@QU[TϚh YDMbpΚp1sp:tWQ9]^cAOkrT)C_SgkJcLl`28@r!@),ȒpLWs ^H"<*Gضc8ˊ:Kq$0l11Kioi4S! v MiYHo#;Is*`ұ;q#B ҏK?T @ƒP:uBTf !27:- Hu,=K׮ IITN\rUV8 bQ_i )DSF\KlõA˫8=:mq2k[0Ȯ\1$fڍ.7wF mȇ t߿Md+)%it|NHV^k(c ݿ-{6$gV,&°*eyҲO2a;N)%sӦɚ[|~4gvl$ V`l +͉@& #VMڵ ~r~+v9bBIҥKXlq7hЀ}Y|۶mC$3Y`"5~T W(M%V4{FD1M"ʧ-E}4k&FSRhSfʖdfrU֬nݺq:2sw;v!gʼnj6|:'ӡe qC#"R,V6],6S֬$%i\8oƎMmMdJ; U3,QAk=-^Ϭw҄Ru=vzR<^Ab+ D=ƴ;-ѯl6t@{gՅYwwע/P@)ŊCk\6+;:]6&y͝;ws[gϞH{@4SLsĶ3^KR^8NiӒl5zR!sdW|j ?{1yˀ1mb*-nxiѢVb3'Ll̘C1jHnjDw4촙؈}Nx d.)%KyɽC)f5^p 81D9%ٟ( ~c%" - %!CF;7 K;#~ֆkS>~Si 86Jal&0yI(зҥڻwiҘ7Z"'5Μ "]4A&MQ:AtOm4s=Gc`ԠAt@@˗TZhtf*8h|[s!g<2~%LeC*]HXX+l-uϔ|RP>@;GKHp9 Up?JCw%@V D;Kì$* Jdt/4*tr'L='`o^Wo&"2>r[$DNͭ0CyVtDTb27ʽL(Cψx߯}-$C7E5cн{ZǏ ٳ|g˖ϴjUkB~Bʾ w"4~#DXԔ`*K# r?e$$#UOpF$,t/ p.> + )FAG<]fl\m[d +)0,ǡ~6Vom!m^>VD39cw H G ͼ6 31ug;>bo>ޡZ4@sH0 އ 8Ƚ8NqTj>B%}RZ& ̪Ô9-$׌Z#Qr#={D$) D},bܬUcz.A_ĸ]fMNwx nI z0!`^jZ5?QLNQ@ 1Ջ`0bB5ʔ1PeQ({nsҸ"װ$r|+wʥK[nLrԀ%ƽvk_Ɔ3/Qvםܿ|bJVH6w\eO܈;{ <6 MfFQG^snɔιѾM20a(0bN8'+-[X MQI>d0Ċ+F, +)JYQ^ Vl#jo޼iH{)ܞ%J (>}4I[Oj83@2\/ǟLj%aW0SNl9ǼYӦMqoHFbeD7#(͛cK0 8q#qB| qE,$6h?zڵ`Ri"ɽlYvDR^=~̑= Go/ $m&OGuDHϐ1cm:"hgbE=,`޽W:;QNMP{ۏ>= .2#!e'\Zf7'Wk+*R)e̳z汨DŽ2Bp!P"5@{B3{ ~@p@5*jle\be߼fF<2HɀZI*^Pꁲo5`o®]I~:S0yH 5)_ΙLj830|aӘ,fhŪ眼Ţ2ǖw`cH'"hS T(5kV%26q2fbmLx<4H6KxsTs@,6>ySDC}4i?QR)+Kd 26m]ʚbtF…%(i2ki~׍.[$r|vbYц 0$/#oѭ36˯B߄JN~?ɫg5Eg7y0_Zo6lim^Yq1qyoĒ )!Z>VeU<_ w}g :b4B`N~` V==m8g#pY|5ͧI@:; p3~ ;cƲHI O]ѢE)Nh琤);:(رb X],dê4.#.G,,Y|ھ]9DYޝt:ba}MŸ>}Q Cf~ʔ?ͻB1Y=fuYP_MEDQז^EKU+^12{S}zr{Z]hKS-n^Sf\CBF.YcagءĢ wp?_%j%|8'as tޝc(qG(`TFC43Ҁ6+cE'! ՜)'GnhOΝQ-wwo{:wURl+4x"D RS25Ԥ )#c%Cߢ'IW"?ָqnHt]~¦M{g2mo̵u)k:x囅U2݊\27~7`b [UpaDaJ7-Җ/7/f"v#G<tX0 eǖܦN?\SEp&//= v^jK5';BU;E;*Kr[|,bvM޹yZ$d =h7C~aNִvp@xsy3dȘ#]#Ax"-.c蚔3[%'CsDEx콆P 56Bx>UK;Bv6HGn ۴k1dH+xzR'5{?ZF=TeH &Eٲ԰^HkU 9dϽY=uJ k/qM[,a˗+Qi>C5 AC(rPRl\$&y~ Ozv!U{}ɐ&E@pJ㋚Lz萪q|į^کb\J*?B[d0wep!>w.AEkWDPvYB=yCmH3a.|;P+0>R߳y_jA=U4ZZ]gvFšnP^-[y%-Z\я?hO]əv+7i$H$m0ebR&y91ACGKپ'#$-]Rl2g.b!#] aC0 AعYX GzNXd _U :Vu EKF&lnh&09'jL%G2O;b 9kk^T1Bd*B˗/,ǀHT  K'@}^QX6iZP u,TI.΃:$މfhؐ$R>ufܓAo PjGaaC7䄟7fjY:ޭ_̚& z ] x,/0bH$9a`Zu >׽ky8=x/aDl[)l$>q1*HO8_ew!σO|0_!0Hnweʼtť" *t=K]TG5vETwUURvs >df[1 XWVJX-t%LYt)Q8{ 1frsB?\JpueS9?4{ȞjLA~͠\cy}~|>{:~{kHqS AJNLmÆE` P q QQ;ߒJزz-Ap:WL1Ћ!KBST!q5?o8Qs~Ǿ#n'w+ RXbTuIb_PcAHlCuBjϭZZfM|QRB/ 8VېDBוn5q=Es*!3o_+Szywo~9>=?12|R ܥkW2qᘠK)qÐ!%.]TC00uz:dg"፤$n7\ӧ~;$ώL^ Ctd ƱW4#tk7R& s:wCh$ϯD ZF b;fd0 ܹ`q=u6i%~_! |d־Jc|/hƨ评28P-Ź3wSlLŤS] jQ}WMڷ>E1Wܼ0[ 9xog}n+ ˃(i6:f?x#uE"l3MR+10;H͒P#M쿩[LV"{!#$Tv`9̙گi~1-~A"s_sBk Gh&kgGH9ww#/^ɯ獹 + )7["akPEo+!!LDT6\{.jJ˩xp U_BP2Cr_ 2a ԥ)f֧6MIA3n\իA,o6 vbҙTj4q NE[a^3ʀE{":x ᒋҨP}l+t- yw$D׮0-{ ӊwI=:3)#PxC3d^K 6PFD9rc bL"Tt""p=g4xWX< _}ۑ&4CUAd%AD)^‰IĤOam\C%^Y V='M`G|MĬYԾLDfX!|r6Į}# 'ԩ'A*a GFbU:p>,;&RJԬoN5IGD5-x|MU (@:.WzL( Fҡ~ЙO???ՠK$ 90"<6yu˺u I]bB~}ݠAl RDdl¢r mD*Wf 2qJ~@8zT NC- iFJ8L|xDJWiGZϾny.t)b`'7l#B|s[jܹ!f/% ˭[`afHӈ&k\^堞B#}5`υ W*EnuurQEȁѻZhiΈe/\flxq6*̳ϽFrxw'9WiҸNzAUdک'Dw#[8R[ߣՖ}I>u~,Sf…/BDH0m9㕣DgRGHrlZIYj TC„)JIZ's_G`c#1=^޽a.%a逧p\5kPK,aǏOviP%=:&ڿ1_(, W-n7fP rK>JƐ=IfB f-HߔC4qzG!tku*TsHoƌ7bƑE2UIq1B x BT1 &T &Ұ:Ԋc/~w4,"cYnÇSgS͙&lՂKeeHk$=4Ƭ7#tWzQFil#Z40k$9k ?&ʷъ~8)I*,yws I,fI)Edӕ15x(`4|rg/Ņqn;+Vsng5R/;LOeS":իb/&DXlE`C3n0t̳m9s\#Ii@[, RS&M< OBx_H墡.ؘ;A߹SVSǙ׾H8WLr?lƏ7HX+(Kv657lАF9Jj`v>fnDW _QxQ ",B̤mh&MCar3y-^:gX'A(rEFBkIH5[Y&ؾnmDPژ=;G7Қ|gWlNfar ۍU}}{ǐ?Ŗ,!&OY!yBtcL8VWH\xi̙}x*EMj ]\!e2$|&Mظ t=0C e&SL*`YXZ%+ $<7?mCJ@:N#kDEVCJdvbi U(]mU>|8B%(R/axOAZ4OhQqV@@0LB#LT4N6N*pXDX`(P'b,ݳ"G!!EBׯڷ/֭5k0IDr86tÍ"$1\?MY|b|[VͰkw6ZTW^<*ײE|>s%`^^>6!!'M2`ǰ0{"O[H)IS&* |DL1={L\Hc)͊Ͱs 9JH]XF_Nw $z3W!Cg\/6nEف]y^,R(Ydmcla 1lX<d~1<,HM H7VId 8$A#`(XsG+uN>ml:taOy#F:FrN/la.#S*$S*' C?J=JƮ;񞾋tS)M(;uwUKhݺ(A'9ţ!.QWrXQ]<[Rq^\>M@O4403ӵn]fĈ8uYqTvbՔ^~O*5xWQxD^j׎JfӇ-K|BXiN]9( %JXrysb@.M6$ְaCJz=BqcA˘1c>g?W/-ׯwa)b1~d[H482pl! [諾@ط;7c )tV$͛oD,!,b"+iBېbNBe1{J@8 D[ )QFBXņE| =x4U!2c!rYM&ES=B@;_HjĺHCC9wo̸}:8lFʱ46)"f `3I&ӧG׼S!)PU5Jz|y]UR>a;x|(+ЫJ59RgnDG%=׵%m I٧+Vl~{pAHtǣ4Ν;_tEqYfiaO| h!I @M00hAFšp:S1vyLBy˕'U*5E.!L3Vc8x 'pimX~:gNu[3+ܯPǼÆ n Fd X0 GAaA&On R\A'ȄH|dr" r 7 6]B,*oVs'ْUNM%s]@]2qGS.[*v;a䴤)9dGXd_?]qy]'0ȢOBx~^^P[>Td5u7<ڰ/p60S)W{FSsPPm㋊1t 9(?da~9X`!C(j+FElv P>qԩqAoqDoj>!CR,Xf  56@)ҤtV-U9Z<"u **THObw ,,6s8PuPllO7=`i1"qr7o&# ' aɓow g$WR鉚P_ s*-5#ټyZPG&glP5 پqJ'»|kM5"7)/Vlv˕)to{C+BK&ÐA51l8ڵnZ8pトWߘʼn2ESFW‰R-@TC4 n] T˟_Ej0B\yF|IqE+( qy(¿RLvyɓJ_c + SYIM%xY(FK]v)EQ-Zr|ݺuvTϰa}w\ 5xqY'I HΜm4Ūi$ e'D9|_ ]Hr"b'!iSHt?C: a$^[J#;֍KK\G w ֬yѣoҥ aŋp)W]!$KO 6[Y u~=Bm~P>Cס݄RV'33WٶM}`<}ZW;l_=KHozvӯĴKko,\0V:eV%.a[]6O[6T{Xw ډ{iYΦUUkMQe!hv6nAafg9JQ!o QpƍQ=IS)&K4ˊ+u?% 쟫^R $eˠcZ1 /MF2KjIz#H;9նm.Q,*r:R{+^hŕlFb@ܞZV$L6-k:ə;f9,64lRժϝ4J&S:A\Zy2FSc>4RC/uj Gu.)|UWf ٹ=ai^:ر)Z0\vn#^fÇҭox='=-jVL/O/o*eJsV?s;b%Kqɕ)<#РB\\l)Bt?&7-qE·Kžy&59~ԩiX{u? PW][IB5ǟ P1' 7HALo` Ӈlk0F=&?'3}Ѕ'OTG/,EHƌ 68*>bٓi@o8KgAFeW&改 p='iOɹ4hРI&GD|աbvE*w˫VNj?IܼsO)FH+JiIJ':ua8{{P!x ]Q_|1p`l_0aK^r^j}1<eJ_AnrIXՖjAڭ͉L͎+<Mr=Btt:2f2;j,½b~N85?=k֬GDҼ!K`λTzx;F^={:Fx˙Zn ˟l &)SKKB<Թ;n~BDd5D `9mtK;׵nOEv[Y!]:Q-?Ŵ03<,6Vmo’v,͘zF'OVa#]Kr.͞=u:>)F#d 7P>GclMK1?xZ6ϳUA &D3ˋj*͙)F<|B!Ņ܇5ฃNBm!R+h둪 1$LiEh={VZ?Oy<IǕ3^,RQ2Fp3F$bb8AY$Qn`6/_^A @|h0#URV֧-$䀁 ?$M- 4k@g-$8X6PذW :^*`ڵ"?Uw1Ԡ/$mtr-D=BVqV(u C~أVDLk\W:(.PuWЪaÚ$RJ 'Lؼxf =6my&M k4_5`˦†B .ǧV3 ˗ZevC۽elbg ܘǑ#QMCXELΣ];ja1lM`b9gW\9!5a6=2gGG玲qd4D?N+l~djD/^vO|} , BA9caW17DǤ'ɃLRJeFi qChb"u ٳj(A,Z$jb} d%:j`cQ9LJlx%ɆF׃fD̳Xj⓴dFm`lAsC*2םeUjsҸqr]]sl|_V;_KF#Df}^Bm!{7|6,G8#dwwVd(AMS~E.di)$N$N8Pǎ.C4o6n\;Qm4:{~~<=^9xpJO=P+8T3+黮EhvM[J7Gӧ(ۘ77%uS'*[Eڬ}إv/&yR ]VJD sY̹$6n\GmYkN_W:";Fo srdG 8].G%D4wJdOEWw:i\GGJ%HY)R1(UuH7oƆQ`rIa*}Y~N>߾};{URvzMHW^:ԫXTV*q@m$.C4O gqA$v3A.^TAbg`7u2g>PS( +8jԨ{wwΝ?.G+i ~مMfV22,[#W>Z ;u 8yC bM:i-K~'2fH\3 %;_fĥ3rQ |4٪ EitKNu̙ׯ#$69 )ws$jgw!F(۩e 9W W4=O<~N?49IaDfdkqu{빊dNJB5j,  ;xgVg BDze+T9s)+Pd]f]f3INxƕZVD#ݱÅ7P0aʀ.iFΗp̙+/_>Ǣd˜ɓ-/ ٻW]\D׊άZfSeCoy۝f@oK& ټ;f>lwW򠢏ݸтɘQ>uSr1\k.zv~O:t 1l/7c3vE4@qeY bG/竁\̼1 Nt",Yl$TK3לΊCnBnLy<8oDv1'"R1ڀ4(TA ->`<c_0S08s, F*-;qb#ڞ{ _صkuZi8E-Gd: sZVD~H Q_ PcY24Rpdu6 eeF&|6bP~{Y b׍ss2duB 9R]˚J!$3a̻ŪCLDNO 4 L#eW%! ^;H]wS aɒ%dX`AiӯYeCrtlW$3C?_M\CgBP- Ċ?ϡ,|Y$4%qi8dC)jL=ri0_0H*jO,'*<^>ɾdA 5c*YQ$!P'3bE#c]BŠ UU{TVE\OФtZiK D#Hג21xp`'E}%p<= 4F$Q%Pw@ M^3/^IReL>|pT>*`,ja$9#D韄l#)o8qcODia]k 4bNEb 1CI_3`X?d`ׯQ,Ya+n͛!J h'q]#۷o \p]%jHaİ^|Ty3YY_?[u?X2o+r䮏Qi.K9$6/ :w.\#~1Ҡ)cP2yb%.;bǵ]Uzۦw={F)S+M<)~Aw -+qz'bi8Yi2e$XbM6t !l^[re(#K:C.C}Ҙ!)AJU~zrF AwD{{Ntt?a0Ra_iP?J g!/֪y'8*:~w,̓E9~M}CrF''!a8~U0?KR@&bi Ҹ ®՞'Nܴx޽)A LLbH",/lj'ڿ\'{V2q[ W}΄!$ oެ!T~PM,|$Qw q4;ׯגΝrV{G;!E޿WP@ѝ)lsvU.=El?}}MͫW"4j5ٳXO(Jn)yd-Zx(ߤyT/ 4W<.zj$ѳ~7 6GL"4! ^Llc"p^Lm2 sOB_Lcx]63{)Vpir~qQt.Yq99@”J]pBJuJ!Fʡ;4'rÆ N<N$j8sTRHpLFҟ, + ݖŅ ;rD'_HE1Gs;OuЊQb& ⵡ&͛/JDyNtX4q{N[.A5J7(䓸0pNf҂)NGu7zUvެj:J-ېP騋۞9tvu$8E7CRNQlH#ś=4N/%0R+ ңdY z,)Qǫǘ({Z͉bb>"hRƢƒ "l.]&ZդKv di0=V]PA* C.`QZݫܸmA"g6EkЀx4E\ xEʑ^?nM5cj@DL$HzdU6V#%#DOCv 鶸_ǎP ʘǾ3etoYk?Jpꉣ={H54 ɨQ,6g0'GJ`OaIfL6qqY*b{gsJc$BvPA 6Il!QD i$KIneECu l޽=$`Y_RKdS!ňҴH@EN׏g/ȥX+)";5v/9XX&ƍ6mɓ|nƌ{NmՅ3*fT?j 21d* B  v-*ۈJZܲŅM푨w($6|ˠ ÷iSz;&DxbŜ3J)9YH'bxA"E|?F玺7lxx0m_#mS/]/  Qѡq 2x0 WIܻCd?q].q'ޒ8cV-ٸqd~*/ĉ,%ؿk ; ;9w YJ +[\٪S+ at: ER+  ?>BJ+HC +C'S+;IBĥb aGJ>u#h6B̕_7qkǏk8]CxA<=U*EMfbo$$hF*Ⱦ…>R=:u pGB+W{K<;un{w=*{vX8V-]φ kGXOϞu]ާtѣcW=R 4K`| &⸋Wǎ 5),>JOMKϜ+j0.CS}ȑnO4G`;w.}o߆lBħPw؁. 4cJ|\P-QV'ߔ :Ղ&ZoU|7=v, 0]7R4hq@xXi! `9 [|>}U:VHjgVoeJ %S̬J&ɲYVȚ KҥDNgr;Sn1¦06ab#DS|;k|;[ w.$f bBa:kN+SϡǨx6F[ckŚ@[=*z s%x󗯨ٱ}Ԩ r@էB@x4b>Z=}lhӧ >z*]7z:%4"Nm-4>R0RTΓG!WC'y9stTL'\sN,QLarYa(N#}F7oڧL+_/$(a0Iv nIJh"Da3Ҙ[9A|)VZKTX\%L8Qdx!&)zjkh͚5˜/3>*V2bsϏY.EL)Rb,m6iOgϰt=dϟI{CuRWSę|{O8xצai.sO3&@B`vn`EkK7o{6mIu;8<@,&>=ű_R4Jdl@j9ZHhdJ!eG`bb!`,!C'.,5 G"}M__{M(mnl̜ķojZO9B$MXD,)Bh9^#xØ^2f6/\0-7vZN'nSx+WL 뒡(6.D; ]/~cΝIuC?2.?n@f2Q"NLyWtqi9{a[?fI ۸ɭ |U8LDž:,D)9ɎGs+Xf+5F2.XE+H}؋%N)ꍦI B=/]tAv+- &Z w)ԋГ"#.m.gl=O ~nXtK U_ը=WBRķ=04FzV]٥!mZ׉qF\؈-ND,%dRtSD-qo;% j*=۴Νaf͊MБH؜aj;YPJ=c¯|87JBu / .г,'V(0Xb?bР eٲ(eЭʞ lxƓ:Ck5k֝9s&hPmۖ/L-)ٳgw֮]ۯ_gn|ޫU+z876lgJZaH[:AB$rɎ!]\>'Ú]XiX.{l)Jw}㏈`p=uֶm3d(Pl X ٷ/f.wN\ڭ[7ݍi$g5jx/ goW%q U3ñ̠+Ѯ^ C\7቉7R6i4CL^^@/.֫$8ȡˇ0֕/ _\2>,h׀ߓ' 5c(^[k0@2W,m$t;ĜH= ̙,vނ=2d3AHe1]paW$,rDA0tUn<g^5ٷÆSʒeӧI#C%K̙GTϞ]ps{>QdȐMǨaNCaMv^^C,Kd->RD{<'O1V WB od?S`n$bj+_ پxYL_1b?`3g!ѫ{xs }id|-X0`ZmȾĂ?bHXӁ4{wҧX40jTe-{pQ[0wb{tK}y|df K`3k֬_C5InUEF~v1Ro&._>c%aKR端"`<֯ `ֱc,&QӋŇ|&\rImɓ)@` ͠zV_/XΝVn$66j]Ʀ8#-Xȃ 8͛"#'}zrY:Ǯ][tl7DvЁQ,v" 54z~r~'BG4FBD~ڣB.'@ԩSkժ/4}G0M78;!̽G3|xoU8_,b N30~Bke˲Q;F!#}ʡ޵!_3 yŜ֪UCJС@ :{bZ$V4 ǣdip!y}+ 2%:u&Ae\D~-2up/XN3b^p >~f='J`.Z,`N"h{$Ht+xׯ+laV!%xa`Dhk\73gN_EgɒÞݒ nȴn*-Ѳ/1CG#Axص~zV">[[#OXr.ِ}+u~ut 1*xzzqb;O:Df]{G簬O-~l;׽[=RԲ@?u3ۙ dNna@>O[;72l&OϨ]/h[شi?G 7_v!G"1{,Y|BdJ*5jxaܳbn_7 e'9a1-*z}_6mZ:}PU*m皘ͱHYTbBh ޿9Փ4IdejVHҗ)ٶ͐7/ˡŽgOuύd OġV qYWf ׇW"M;g5_IDAT_'|&CLƾF\ְFƂ㋌M\_*@wEn$!%<{7)A2ܤX3dyS,]ڇؕzen=2BeϞchLsi`$̪8@z$郄9]ìmH+ 8KcؿsH;u֬HBjQf6W[-D`0IbB< 4)Fݪ>"Ə  D1&!_WRiF'[' {Wr׭udw!.ɧS{bf!xs~X_ƞ{$\% T=J)F6>Cڵ0V }Ѣ|9g&غ8V)z K"8߉am AE:?,0J_ ^6m JD7`6'<!(bH@&'O8@8-1"ĉ 8_pv,y.=3dQVFi$)Gr)? 肝 WXjn<_oe'נA'O#Bƀu!+"hފ~Z1B?|!<=y Rؿϓ64{crVUҤ?V؞I@~0KDP I0d#`Ȟ]s,į_~#lFm8#)/"A'O|%4i"z$|թ&'"UdH#BQ.<:@xb^d} }cLxNpcAK֯wZɪ7l;6Fݾ}BתUD|:q1l06}|ĸ@3U A2e2 $v;w(8 (euKtI͞D'OaȅHbQG<>N 9kV8>[ O\aߒ `?>7C҇v\e˖597ּ{drlq;TBl yW>4xbؠT3V'~Qǜ|mͮB{I}]uhu$ ËƸ^>u$':9r-d7opڟr1dcךa"@\KX[YBaK1g} > ߯zJ1mKL 0F/fiBףGn;v8cd-'uJAL0=Jw ! M64*W^\;p!b]E׌}_l&v8or0xx".W5} 3?T)_|~ (hu\gਔFٿߣsgxxI{6u d̟_|n*˟9Ox"3᧟@R5 A5|1Z"|ATRLaZ +2 GHDs{<CBk3Iy_{yy!iK7=l5NMkej 蚘Pڴ6Q[R6eAՖ%M/^?g|{jKnn$x͊ i|R/ g`)<1IZK&1Mo#k:ph(nN{q-'ܹb!o$9Q"Y@?c/, 1S!$ >B"sG8~B4g61PRÜ|o旟P.YՊ$;߽0V|JcK}ټ9/tkf{kלx:OL+E1f*ǣH3x=أsU(s+ Đ&(C `!!޲!iBWHƇ@bŊe˖M4'N'HW0 DʆgVdE׫GHb%G>S_$)2T"e7ˆ D5F1TOɷ *Z;X#6;_~百__Kս{x2|Tf!1g{NҀ4KlYe/7PہL+uQ3޽#3f4PچU tsT?f@pK&MytiͻQgG -̠4ir~W7<`.G^rڨ-G1E(n[kH4De'<4f(c_]h}Rw6.>Z/VLs7&mdKJlQ=Λ7ާ۠A$TS!1٥g*s$DadjqE*BB k׭ØcW>|8(!ĊjWkrҥ\IiU}C-[@u~F\V;!ú&}"CW$$Z}3XDV,Kֳ.Vz%BS8EՍ,eNVE-+ւx\qµk><Ib] %;\e`cYe a9Lahx"_mEsymx'2"F",B:O;)&77bCވ>8Q;U:u!j4}޽ɗmP _,H(Xʐ0d| nݢ$&lDt.\PQr%.ۋN2Ѡ3*(h` _~U۷nn#k(Pdp@KzY柅 .]+:u.GB) ղd"N#(,A'T,06P*NdV6_HCT#Qn7lP,y}!Ο* ')GvTבl۷S#JufDA {ޖYC֯_}}y:-ƨӥ9r',+5jxh3olX`ɾLs>$؆ ''^f>DžK(! h `|gx944Ʊ~a;};TC $;٣e΍Ѵ5˄%alO:]_e5FeM&{yv2-i^m~[ #,g, s2ʆ*9ƓE,8h jB Ӑ/S'5Aj%ORϒ0R*h^BAm8nHfE# p8P†܀/tҥqAnE"vo…Z]$Qb:a@c¡2 |.Q~!|kvv80]`e*jK~TvBFIPӑI*kMԖńwՏuȐo/ªk(;Ҹٟb`rSUQ@q^5DTumf2bр9gm16¶5%CɈز9%\TL/Q$Z4h@7ʃlݩ^MtARӈ!/DGgVJڛńPܖFV=\qA-<axoGw֬9rj{W fD0eH60`x9!Yg֨NoZ)Zl)U+EuT+njL5 m BRH3fWXF44ˣT!kZP_d/4)UZb㘀7A̿65\7Ku氘OJdgҵkG\j-uRNN6xfNbpҥ$bIV- U{`fA1h5VFݢC14Sl*=Oǎ: A8rvd[qSW( aN "R4ӧ}Ys(+W>ٰzXغ~~sd9@NxyNPS$fgQ@WdIuzYacIF/EJ+ hV;_FޮUKzuHw^*o҆@$C"v[-V\8%+j}Xm9'y{aܺ Y{(Q\]fm9sΛ7O?n+FDͷ_xq s^9ZUE߲ߺu Ood=wp9ide׮]vVTGjBkx8z+zZߡBⷝŒCk4D+b#<16<b_RQ0J)wЊ5/)1rb8/];4oFݠoŃ. .#|ӈѼ!tnٳ],@&Q_Ѫ'^HÜ`bҹ"%dˠAcFA*_>#,%۷ǐn!cqr.K?7i0H: thPE6\]QT2ggx"2Zh:DmS&~,5 ) ]t1Qҙ2d0 ~@}L>x g>FT;f" ұ$WC혷C<>c*p 1e V iy_cZ2(#Ԟ/=L"ل4OɠJB6Kc>s1 ܛhص% W1pfß};Y"@tw\i2I[8TAWjEl4#ttLN,>z`2E3JhcbD |ƋM.nӾؤ{^%Kqjj  ;}۴@a]DxtTމj ~f'A~:_Rl^xAw w =L)4i_ޯ_?:I>=*tr'=zX/@,[f1XijqNP`;Zmj}"̤p# a /"Etu뢞FPln7742|yC )3Q.W/I"8fԌaB|H;" y/vo1 I i"MU\o'ْ3lٌn*;)df2,:hu'Ojp!,aisC۽;dx`%ȁ*O?K)}%`E4w+ fy;?̥^̗Wk뗰0Sfz2ﶨ,& 0F@ɀd<L7Ӎf|z/-}ٳgLyYnDT)4E4I;k$+UJ})r#iEMf=:h[BGB7BaI$4HVMO8\GZ ws?ڥu(6mJ50G&+~P!k>}j`#̉5ֶm[Lv牓!hwȑ#SoN"#v e7q;$ŎU0W(0k@cٲ?s^,6_߈CkW5]u!@F NB 3yr i *a@D`/s$28q'5=M"o_Ŭ 3'1_\cY@gf pH_:UC'K7Dqa cYqm4ǃ'1,zC4Ăׅ0@#02^yH2#(3'jY0| eXNwP'Ў.\ְa)L z j+.iA6N}#:I1n!kr&t~"%aÙTbQ*1c,2fLz<]GfJL* d2 v>13YSs[$X+[=(Rd/ƊQ8 w;̙v.~9ի5+VP˗GT.IY܇,zA·\*hlm@;ǖ(O ""HM6_ra7TNi0۠A/_UWN*Ic4}'^٦KR5W Zl}}4qF`#cgr07TF +@øe͓#΀1(x3*F#FǶJGÇտ*pm?0foH<@/k=GIZg+< 4w{ICIN6I* 1B3Υ_AÒ%t?Wn]"$YCйEs瞈ÇO;wA%JYL[]%6XcbYȁd8+r&0R!̥%] Y%VGN`Tj|*eE}T%<h`+'JK"U6lXx}E}ݻ3S[~Z؃K q?*01 y^‹J5+k/Vᄭvƍܹs;t#VEq*ƫOa$$ԁb:SuuĈ}lٳge˖ l@ZfƍI#C)Vet}<$/~r.,W&F%zSJ IpK,`浣[/TC ҇pZƌlg<994$KQ׺uĈk}*U4i&aEQ/fnr?RLW7%T`cb5cNՠcZm7c UT!/8!%Uߒf~"NH)k H4seu +SO?4iҤ|"4| Z/N0vϜ:8_TI)[ 'I/Xt)"/LX3gZQآ!6.|R{%t;f)`A(em"("zW;NܲZ]}btjPEUC=QzΞxV c \va<4߿P(گmaV3gN6'&)5Rۡ.rC14-T=2%e˲U+! =&!RVŽύ3\ȄͣR+0e L[p1pYYoh(/ ✌+flR4وck <]6}[/]\qOۖ,9?)g*" Q(t:zT-d<#+VE.OX93K,^j;Dh8 qN^P8WL.. #C,`qCƍw!U+nYNmu(z\ԫ ,%f1r! ͳ}{ !GIѐ#W\%&a6.x?6|st1ߧ ~M(1, G(_mNU6STA$q!HfԨQP|0 Ҹ"Jz|󦒚f Ps{X$IH^.)o:4 M[/R9.@T\D,'hʙRfY|J͏ө5*Vbrb8 "9-eJ1h,""X3Tͳ( >hXKJ c58 gďr|L˸]D 岭Lb%JT #P#cƍZU+  fA} ue܇-xK9FuJʳCӺ5`"V =H.%Xwr-C Y8jddU-Q1޹R~` %)qIDҷLl@sdV_ 2q7FR%xw>9 $xGT.α~x-X+}H2`N>vtWAD[?O°ؾ1򘳐+㳪UD3GԐh{ĶOY=5Ex̼1_:o!kWQC̾WC_l5~獯+e^k҅g~P˹e%Y@@E]A)"E+T]&}&Vi!HdP]2z?ON7nlݺ5ؐj IC\> Z>,BrIf l5b;q3Jeq?MC@oB-wEgh4wnZ)i[-˦}&{~2a;HGvPYea,1 0xaU5"!='QBSB81B&'KI ejmU% MYX~(T%+ H[:,6gK%~!r1E6vlˀZ9z}j۶m]Q&_|aݱ3-0„ nmFi.3FZ!:Y9`%믮ƑaBݶ8EnlSWd͔Ym+خH{/:l*y)4e92T)OCY3;\:p%\vR":E)d5GFhmx $"wi.;Go>֒YD&7yG3H@Hc}dMgy:xԥKia/hr'FHhI[ kt U miu>,V+Ϧܟd -/\h6ݻwZ \817pd>T7lJKB` uҹ! ÇƢG Bf_@6Yzp1 pI.wY^zՉij;Ei H̳ V/q9nA߳~(]]1P9ȩ\s&_ ͞1GB * y% |Bp4ťIt* Em[*N96b]#}>Ӣ 4&>þdž'{AKcc +ǣ itF}]:Jr |Igf PMNv_zq%N,3G m5"FF>ify={J-cBmy?0Ou*6MP]<|,PH;xԨX9/|$Ȩ:z[$(L0*zDLɈ$ 0T1ɔ8zpׯ_# Ҟ 4RԘ%! |ΝJVAr@ȟs?s3!|tf?/KJնz.rI&E|aoRTq0H*L]_ފYs=2?(,e0mLL=Ƭ1}z2;砠}{u-z\RO7q_SH`&O7_ڒbaF R,ͼ8#v&Nh/IKf Ȅz9E"]*bR%)VcW`;h9_>~l"]^);MK^- tć #Uc=_kR%B*&Ρ 6ICetGt(!?O~㒫K\ Vp{_"вk޹sЖ-Zj+eX~xyC<7rpHbuխ@ Ӗ \HC&L1VHtɓ]IR'| Ź B5B_ 1{JӆI# _ c 7i 1O(6h.>m o_T5տb8S"XAyRcI j>Y~H s/Mw߹f~#&1ЭCRlYc3M 6y{CO`%R'O}:zsbeg/YbA=|?fjIXju „ ^ ˑDȈI l0Jq v3de =2lɰk2FY1עs`2'}^{e'd}gS܏zvǫ\?Z |V$Z'eb>xI$}? $՚4[y=Vл|pR f>?m v `q=ȘĢ6'S#%ʅ<2BVZ.DvBy~MCh4Hdo!Tڵ#G󰪊dt:TrS5LBQ!V#t%KVfoիWH|ԹG(ٳgǘCP.uhM62pCLU(E8ҥQrEQ&}H byF>$khp!]+szȩ?dwtg\Y?_J !OjS췎|VTIpvb7\)@՘gmyl1d.u7eb|bj&M  Y\⾅RUP g=XN k2!֥72jDjS=0tƋϪ+nkq};?pYdYSE *+n\8+T/ תH27MW13QMy]fDGk:ᇤzI=֍LS7ˡܘ~ Ctߏ}, 7X\bP)ĺhR? 8Aq!F49rn,|c1 )UKt˙ݰ$nnʶNA|?|ayĴ$䖠>GqEa10"F2D U eQk=.+W/&= 3O'ۉ)czhHn?߮*)BzyUaM|}vt`E@._69sOdFrxg!F OP2]ɨUVcQg86q=o4!ۿP6Y+  ^e!d7 !;ca }J~R}ܿDHM 5ldmLÇo |#5ʠg2O۠\χ4&II)_|R͚v];j9he0Dݻwk]쵑 VY5ɖWPC맟i (Q"q+exRKs]}Jif۬ʣG52aΚSXJ$]L^blN#&:4byW5?0'imu͖x`;D`c!)/>)|7O(A M! [3|{6˖Yp ,_  VPGIPjqiy`|SY%6.!m=RQۋ&SMNԲr+]TvP8u?t=AWcz2oXD _}҃xBq\I$y#k%L[B|fv#H8jNj)H(^H_Y [/!C*R)\IeN@LU͟w„ l@ǏozVE̴HոDk`wpxb!PB;9"hEW'rcF+Րr1`ɉ&CDUc(@\ _d??{#vG .L A͐0/6 { [x]6^Bzh&ch@+_bWFjtXSKjWY70gNn݋J-/1]eHI350#Ąn 1 ӳ{uRqȰ3:RF_9u]'NĂnV_O2g}s#<#FLOە1u 'k(ME+dW4\O=;j~-YN*]̇PN@u?)MX'8{bwZ5dXa?Xm-RD)b񣠠K[l^"f4.5h~N,5zZG2ϿضoV/fwK$Y[=4!'MrE?}:Y=:IU*ӷni_.DdY2_9-Ng!Vv @XjU΁$wj;sg/lmCF4RаsZ]8aՍ~Xk!ԥg^O)M&K푥QSOV |7v`|w^A3$a5jY*W@~6rc/poBЀK?wm2yrA B""/_AT~;~9@/|v*ߟNPex{X#!P{BQ"$H&gY#B\BADom|&{Ϟp=w.?o=^+K-|B _е׳V%y+nIr ?2DmB)C~c.ZBTpR _I*!CdV6f!?GC4H qDS_JisJƸ2y2\*XB񎕼N׋X96C롎\y9B(j:))494!I? 8$i9tȑǮR! G4;2*F|~)ݽ]2kJi8rG*-UӧycmгM\x!4G\mr8@sz!֒~B!-<\y~mJ!kVu1*Ź\^r[CXĭ_WxeA_?]t8(Umh6 'W4xx"a0m@{:t~UCY~WܰP0IOfـJ$#^ѕ#dE\ٳI\w WN|R:~܎ MP\֬7d{li&M+U jUܼ֭xmpgTbxQZԩp[%D'Q;&OgV'24ySnP/R\U9젣)n2HX-̣J:̀&`O6 75?.^hq/b[{NW>4fP_/[\NCDHKV$Z j)Yn"K E{XA& #]['7?/V^ek˙KzoAuw]ώ d9rT]eDrFAvTEšx dMԳ9K}{R2 JT-c@ ' u?…Zch0 PA͘cju0qcDjrܹE Y86v@!hO3NRSэ2!~tք7Xn0ZCr0wDGA8|?vႊް & 6uuWKz6(lشo,.۷#0w}A%*_2DT`&/4z=e֌JEX-<^yU >8taݐ{E<9pTR0:$.k_6Ԃ&i?m=V .F7}WV$  )1zJ7j͘ ]De)b1 7jEuyldj1^r.P 5LT[5mztB걝vL ĩt7hA(@*Yy XhMz{ٲZOJ Ҡbӥ:j, eM˒]uD)8$ >Ao isΝ;iIQJ0:tOؘR!yGp$IvzD57]+d>0"%/ $V؎Gi3f5:aq$2/XQ76W^($u%a4k^j=Q?ް)܆ɪdG+1v$FJR%f%bgRIPfJ\}>K3vqc#NjRҼPp F &̂&C2&Wl2N\w(+v yfe2Ҁ $ 57{8Ĕ@A2XKy"qpH$-? K|[(`]> F0*G.ރ= /~Ek%. R%gE*‡hWncxGÍ]2mTqIDJiޝ'd&F{|%KHC1i(҃q׽.S*>kF1J-#2^2h?vϜ9slX٣MXKc̀fbwsQ jƍ[|߆<#! 9rpGm҉'lRi2eX:vܯIhZu2Ԛ}{xkiF,4֝{p{˞' v%edD,%(ፔjw7A(+sm}@nkc3LHGNnH(=laaP f<>Bsy RD3Èu:q/4ꉥlAJ(Շ_J?@ׯ`̐c4R :ditЛ&3eU >Oؘ@poDd\J[ڵQA tC ֭ۍ;wz{g͐Ã"_>{}mpGvq"Úm ֶhŖpkT۸Q*=ʁ1,[?<__dBx=F{ɴ=eqg%,_z/*1yD5q/ 3Ŏٲz ?N'nu@*U ቒbOe'j4^K"e#O L}̑ai OwZZxn@#! dגІN2#TnD ٸF p޻ׅۄbMPU\7mݭ[ 0e={EGG+AoxX0[Y,qKNn7+Wd6oF6|SW[|{^% w7peo%h^*_:cl&qm-cr({#`{C}Ř=M.ޚ{u%%b Q/$Y{%$jψ4,-jsi׮=?bWF$bĩS@s`bjeB9 dITJwYLa~dK#U IJ|^*=րS-^,z [1C#a0H3g~)"J;s+VDF>wnq  ~0f(]kOӂ_>q wp4-r8؉9s z@WAuҼ2nt0媛5]͡7„X֚6am"ieaI@@"pv@_/  ֭|<ܩH7Um& 'hws gϞF:t.@C7kWkת0'c:4.ς._Yrc*4cɖ!\6SL)S=O>`:7]Ig}p{c[~[:`&_XBŊyիgd~Il=4E5FA'O"H<Z<6֭Q:so e֌f^@$czJNf3%0C1W!`%KYkyq͉T+I3hp߈@f`|zcfCNtӲR+$pVS 80{gY;B ݭ-PhB "ŊBŝbŊ; $!ijo2tnl>ytsΝ;'5kVz~ Ӡ׏3a@(=_ s< !C[?Z^Ą`ꈄL| Dbre a֭kw[βMLLԙ-%a?Bhۀf)/^\&ܹה(:xk#m'ݻRN1 Vo;2+ M$dmUn_7oOYI)BAEz;io2}'ft/@:Z q/lFra-8DA*!=o jΝcǎT(o&\O>O<͛w< }qIL3Ǐi?cT~ԬpLC68 e-F%dIƬCf[!ݲ@'KLkXǡC|lq2׿bcǠzKkءo3_&O͠Ab'\OVxӃ(h/_n.W׆O\ئME BwI2jXO>qJYfaAM&#JSki Lr!^^sLtK&S)SNKI8H0@me}B(6k8bЯUM*Z8|oDac2Ɠ?v̓U-SCjjQ q$ZO E6>$qxXO%Z۱#hY@%tD*vE-Zdƨ9œ 0I@BOp{3F|{IFANEKʖ?-ɓ*p>21Ai~pې)f*;uX$]&RR4PpmM4,XYv॑ \@j8'N(c[Ҿ`U)Mm$``ՈMXj9ٿ̖[bĞ~J8$5%x4TMV5K~tև{bdd$ۇ;6jlK܋4hG1t'K 'E9͗_B2Qmj!wnmBݮ]EXSmlO=Z[56gWotuKٰ-E>o^zi߇8GTڵ"Y+<_yܞ}fețlB1&%AE*x )({+t=wf{N|ݻ'/Vx۟:̩~ 7`F//L~bӎ>x?]zqQB!xl턄hPIGƴigUqm4q h3Q4߉$ZP )N`hҺ$"Ja-kM{IvvZ VTPGwrNQ73JڲJ{5@AQZyU eh>Chn%'s˖'NJ҈R֫WO\J?\1h4tddpԤ'ȣ@'; S%Cj^G *\.| ĵQ#XI;Pʊ2͌Ѩ{ Cʿ2H) yT-q떨|HX BBEri$-\$_Ϫ 7?0wow}VrpE pLB 4'%̎(a6eh~%HE6=y#15ktޝG|t?&SѸkР%>N4n]tbfQ C頍 E'vvzHwxoKիƸcI+HH[ۈA;S GFGV;{Jq/ -(;Jl\֮=cY}6a8gY%H Li)M {]X CΜ a)bڕ      !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}+R~}m4݉oSG&Vs~M#g;Ѫ1&FM#N]R>SB)[Ν>+X H=#wnOxy=v|~|* "Ŵ}Np":1;}l)vDGf0msY(q:olɕ:9cHl0UUPe׽%Z8̀(8XYLf _;viذAr)b8"""1tZe\p6FBڷmKClNjY_ßLxϷ<魨Ϛ7JMf9w_{:!n̜ͮnV7%KM7~ar#c[Z7&B-,U2hӃ@B7 pז"d=C0BmZh)#4'{8k)I E2MrcoNښfKbE-h@AUV6g1KԨq;|óWDwz2G"m 1Rd&9~Uv$Ȉ)pFQp>2j~W0ZGA'[/\8S+'Ydj=Lʤ4ucO ͹19S&F$Ɓ0vǎXJP#[SfO+c~BB\7(mC[Hzz؍d ,y딡, I S,#+WhAD1 V踟*_1Õ3Hl\A4@.Kٚe\w,z%qԨҥ 5%yk;RZdI4QYu)s(2yr/o$A424f'`NȜQ֬y*Z")VKuq+$uP{/xBk]{YZ|15HY4?^=;9M&&OM=ߣm bK1+jN#L EY9āfj)jNubrbhreAgtkC7MnƎ0/^pa ^xCCQZ] To׹i/jp60@}RE&387ȸ</]HvCaZ2~ߧ~8jNc/-oWu6eKA L }`#4mea`@RJ*K*d JGî6ܲsۿUyh+<{V24tNW2Gd)Nm!C7^c &<mJB)ٳo#cHY?Zf~$u$g}c.yǖ#Bj+ M*d~ G3>Q`2 f=rX\ז-[^D'Z鲬x%7{t{9Qv۶u :ÆiJysۃF3m 73 \)Rn]ҘI1|(wbcp,ZiXgC/|JfQ V^Dh n>E{n}c"n%,)D$(Jp%(MDMHk[god-i *_I?ژm<}- U:͟Pwm|s Kˊ5{{LWfFh6ar<>iT$Kf~VMJ  OGӧ#…b)>A˜`ϋ*C{{͛ (`EIS,ٲ ԸyҀ 7<8|8?U4~Уo񧟒K!(*$Jm||:dYt'Y]gnࣵjg?.N$VTz~\y}A1'DXG: `@jgqĨ=P#?Z_hȞ\yuݻQ{hR@&OM͆e]Te"-&Vc?T_޹+R$f|j8,~38>toutz*+F %wHP2v]gF}6^;d2x4ݵ7~1,\7Lcl8 :Lэ0!H0$ƑMx:vtFDhԪwaRG,oVp\\>ᮟO 0ڏ&ْ`BYfXʿ7GQA޷ο"zi_GnRE&wi"]"P谝Rk}}-,ew_U{j#ؘcq€5s7UzJ[ZuDQ_S߾at2K\ pX#"~mm9  xكK X$'# zb.Ҿ[iK %6ɤhtb+WC1&tNTr}u{kUM0!H8{|"ևdrQ塡vܙ;rÓ.xN2}Dԩq fW"J5YELghEDR(ϋ-x~1!۴v_\i#i}@n-+"qI?ޟ3g:D}-uȌ!n<@w$ߟ{59V9GDzk_J|wkѨe_,yb5k9.ndqt@K VY/gYA@b.4 F%=~cRA 䄘!s22?$yfM˿(.\){ f2D6[*n*[Da,t2 FmA~#I5$d%Jx p҅jkW*ICTxRpܒ5[܆/+1iɪȗ+|Y0{CRsFJD/V7cb'&T~P3Q+䫕-뉮A %'Қb7E;n91ɽx/m[@<:$+1GI=U Mq*t+зc`[fR(iǏ}r?o䣵j""!Y5@MN^ TDcsHG]˖mj 8$$3l8sHX_Sk Zȡ_uSL u~05]&m5t2رlNC3∘R! W_daج8>g&xcjYQ^<'\u|,f\pA۱#lȁ[QƝYVC=zT)mN\C~V'QA3g7ci#̢PFz|U?ˡ!KvDIӺhʕ+?IW@cDREC#"N]ğ@ Ds%3}4G8Rm~7Z5qIuD˶nݢn }ct>\׬ˊ<"xM˖$bAwt耥ڋ O,_ >3Dy<ʔx JhY3|+:440lӅA"eu/^GgT IƄE!L@VhD!&3iy2g?r䚷wK${4 ƇYkW[{e&SڵSvZgNWcݼ+-;O 0Ї$8ip~{5v)Kk/i&&|0矋u͛kQ1RgViP_$`{pe[ɨ6*LJ! r]2NJS#pUD+$מRB1ɭNHİ9̟d6W?s(CbM },q.gCJCfƸlת$p *&CjV$-6$_B67KcIY\_>4KV:;lW9O‡tڱ#\*|]xFVBqx&S9Yئ(6ʖɔ*֑¬CGi1A=鿾W!ʝ{^hP.B 0|wʾ'5%oU8d I6r77.&dݦQ yÜ#yXֆ&cwB{UvxT4JwBܪ-٘4*! SEfλժ8F{M>=Zs'{}x4NRr%?jggg2h }g3qZ5Ʒ[: -Ytb',+x!y|HyItt`#mvhk~ >Fl!9ezSxT^DÇwՖROGC.7"]+!zKn5SA^iژ jt9b.Oj2:ay!d|1dc,qX!ǣĴ<ޝ Xɒwﲆ\vlKYCI6J*Ռ?G8ޞ AVjD3guC?re=;LL&kœ(Hpb0 :~l9C\ݻmv/S'T m*}_8~˽cD˜+3ʹpK{GgO._T͑SW+]wW/pHĭd|y#'\5SJJ!!Q0~[fH &2$W)&?&y^"O"vdc\0-C?˳bg͎kb߀-X%{YUC `TfrϕKLhDH3o>gwmT?:S;g:$a xs$#ydҚM-cj߈kt˲e/RW%^g$evTHxQbiFZ +bLO?ˎG.ݏ[|'MrJ=vS[T S ~Q{Tqscm&8{½t;a>/B+ j.$1׭Or($..oYwXϲ&Ўɤqd^CMSh3,+Rx'On3?(2(3S@0w^_hQ,U(,R{gA~r7 w NsD#3gnc-Cczҽ|V-{v ~]QYGJp1"~0~EDb5=9[WW+?lR\:S%ƜPlV\b`/{P2JV~s 9rm(* ;FAĮw] 6.Y"jJI%7xGJ.;>^F DsxԩJZ6 ٶ{ .TkNƘ13!AaQƉ78w>oauT/7vcFI{Du%׍j m[\=TC pzt:(tb=7`/Ԩ*O kO>נAb1k_~ACSR˕}&($Ik"۸t[vgCBBERvbnt&sRQM)P6_Fdre^*nlx&pLJ9o \>|4CNe.5b2jzۅNtsVk2C럗'Sʝ]䛘׋D{xqoՂϮ^RΗxb`I)F9E70*mSVOTӦZFn1ܹZr25m7VY3ʬXw/ l9V=B1%F4q W`*F_ߍ (|Wi-;zJ8°?N&&?g'7h_!LA45~14e_]d9JɄI#EW5={+*n-R?}zǍuݎyҚ ^ޅq@&@4\[': vł]]8ݾjFcm8e7(zfʹՍ6|X/=^V)ðWV?u4:Oh眃pA[ϟ.BjdoS)睛6x;w7).c;vz۪gt+ڏ&ie(`T*d<Xr.{Dqp^s'K_`e$,Iivs(bvD.1t9vT[ j^!H'bɵoߞ|KJ*Rka=X;qZ5 :1.>74<0YBʕ yB,d]`GZ! TɦI2 ܙY>ܐBWٺgέrʋ'<$sCeD]*{ Duo\AAǫrW~Dɮe"R~my G gyQ]\Q7;lgHD]Yww1O|vOU7M 1S+7jcYޔo'G{WAjAr0 o:vLl`wHkvD {!r4 iYT*"?LCUyGRc9 j~l+WZW7"uF b ժUȗ/_vP0'E"G!> OItT sE żgz~`1>Vk| :P6[/O)l+v,)o17ЙJ A 6$ w +*B;]tGFјҫ4M! )sŠx y;VgŻ悏ayu\Qjyϟ<0ls'8툹 =4 X2<6zDq͚X~#P> kMj7nygUEZ)*#^\=ZR^\u48#f Z̿͊YH0l0uaP~n$f@8.b$ѣA`|߾}dVZFj2Ik !8P !YxH|rpEHs&7U7opm?{8S} I]-c.G$CYc)֡}!}Tٲhc²W/WW*>݆u+[z ޛs)wC,6Ů)lEYNNݝS'  //iŐ6}SHAه4[Vڔ(VXq Es6@D2}]"$ Yop uR3)~¬qNKGL[ֶkѢ5 F"-!}ՉRFzHh ,o\@sS<|ǯEI,Y~z;i{&Y3]l:-8or/7N~ ;ݡe ;zξIkz_V H}.dps k.SLP}m7dEd$n4L6|Cghjqgt 5&|X%mP]H̀#͞{wC# 'PQp!OA(,."W]LUPF- #o@E*JϞRɡߨ$*HW'M*Y:mW !@bl*ؿ`` \e[v;%R#CfaƠ YNc!!^Jzxb}}pG.b<(Vد]KXPTuB-jt/S&v !u l 7^8Wb ́vk4/W4IysپMzem sQ)xF&~JQ'IWNXJm=!ebl֟F)v瞢@b,.rpS&śZxgGyܟLGFxZ z۷"X3d1[1C;4=zS0+WS_`Ş??v5*]\~C ,  ֪u e| (aʷ"PTUk'HCc3G -0ׯyIT R_#GRZt^lk4, &>؜Ok8Gy@Af@V 1ݥ.ZW 6 'A7izIy;ŠJփ#e &i?pZ/Di%EG8˗.m2iΟQA(P%,Se#_#5Wn9yf7VEUB SLyX3.;~ ٔaTx.nB~I,2K7ˋ W턨[L!Hgzm6dLw?R YBu e\d)^ig䊋K񲕎N.A\-9"E ph45B ~3=mfLy ['-rK\VwP0godO ^Yp`{3HԌTڵB!UѲPH(M?P5|6T+x0 ЛsT ]v=L"Gy]ZUؚe+dTBbذt`_Ԣ._I@"at+ۧEKBN$edf m ֥H%.ٜb=E}j΃P&, "'H=b}GPU*Uansf͈4(Oy(EcG󵭟ΑW(E[q,*y$gs?lVhcyx㒡qݣ,|D5z$iQHKNUo2Ř"@%K[ ;V YJ2IUZ8ZذB.,}v2!vΞ-@0 2%R 0WLoDj_E['af9 6j#a>:/O"6\?=s?ܳ}R~DžJ ?W*T=2` m ۛdg¤gOY;)o& ? !i x >{ CnȾ_{Q$w]\T5\Z׮MΥ-0 ' (kn$|%P2;TEs{"KkBBf ]@V ?Gs*PVRݓS k6*Xa${tn-cLhnq51(!rU*ヒ]ʧEgkZ+nW9suziyۊddP?.a6Ӝy=miucXoWlGUc. f՘YkrԬzٹ: 'R}L-G_xhOg8-\G\Sgɿx]%X/L03L6B9t~~9?F K2x {i!&iFC叩`~%0xcM^D4LUCW6%I2DÆ8r61Kt`f5YYSLWܾݲmN;!P (ä;s!\% F^@ѣG=z!d]:;P ֑%Rjt6&QǎI3vr0j &(6Lǎa\a]p7cY+-\hEN,=6R Q;"".靌w|!4$֭ڕA yǓ$Yׯ߆ /f9|5ZI3@~LѢE.ޥӯt/WQf"S&T5;F'YQ͞]:])PgʎӦժ;v̥զ-K'r@4H'̿Ly/ZEM]6<>P(;\1[`4Ib^۾?7@O%.-ZԀbNttkaK7/X$ڶmv&8dwL6:mw!Y˔ovtQh )-eB ]6TrRh-[agL5h*NدpNdaHg\1qfg;xYO!"\"F6ґ@:jժ/t%J%6/Espy}-T);t2a*Um6-VRZ[ ՁFq/QUazؿ0[HeJ>a*@qRB3OիW1$X53 ?rc䵔+Wn5kMߦㆽ|ׅUBF,e?/K\KXKJv}Z  om)C""1?AMɓ弪k6ٙ?HCsEF$'L_P)ן|b14D&t:o]:utz+ !wJG?9ZdB`lV\ A#}<*:N9x@/>ʕi)L,oWt$&rI}%#g5XC`.[qZ\ws#/^V TZʍ9 FyL"td0_43r-#w_׺ȃnfoً}P^=jڅZcijg avlo,ԍ ;&AMU0/#+B8i{oR&+#o4quarO : g ~'~bOߐ?|)1g qEL6kfwE!wuddzpF3n,̦&i%U-h)^X5`O̔ Ɗl@WC 4_x:HDYTXLlזыOMtʷd !Z5[Ʉ\\U/XNd#$2!-L2Jh#؈abX/a'U Q|j.UfϢR!8^y1w^ݬ7r?Y>,{,Uxԡmڟ!kxՈ=\TW|ᆪHqzw)P /kǹ).\!|?}8#v[˕EPd #GV/*oޔ#WA& @|^c/':*bz~'ufl9A}UqK. E#F&>'V 7JNR9F/͗+k{$ \(X'k|V~Eپj;SǵAݎvmŽGB/(&QP`H.'5{PV9$Np+Ѭwbg+F ? &=btoj,st?y=sG^:]T#L ܫհӆ\H8lT64mZtpWT17 -<$ R| q_NNZLd.rz~0iN4l^ݍ<)PCC B](5_ =:Ij Z) S a!<-d3mhfD}RP}7.$y\4 +Gz ⡶F+\\k ESQY(!& -GZAc?' ^(_! G1e1". (@ii y Ѯ*]A62jȪʵgvSG9@{蘟yF*PA#,x^GPP%GB/_&{^hd%?r*[:K֪y{CŪVN=S;"Z{jwrk;yx1ڿpХUp\FciǸq b B⎯_o$V"n!rZHe'N(wSųK^$YHJL<v\6,ϟm!XKcBg8ŕ,dc{t%ԝ뷥*WO:o9MEܹ2e䥥}gU5q _S uxr?|ϳeم<<ɘb]us̝7 c:!@NWRI*%E dJPfH{bq\S߬Syi˪%FbŋP-]$뫈a2UEr(Q&#_X2ß $/_q}`{^vߖGP-mEn|ȡRaܨl?\?NVЋ2.fY1 +7qU(vqFoC莛 |V{966'-;~ U!)[.&*Mwq4vTSBڷ;VY_ Y~Sο/){{f2v-tpn,c`C&8?rz0Z?XO.j'fJ#vHj|Q:Od}+;bՒu.\?zU&@V `A o*rbų*B px:ymz8Jc@;ʚ_5ށCǵO}2HӴZ%d==ךLpY$D%&&&2$0.Z~{lK?nBcQc˂e⇭lDlE0=J:u.CIH;qՃK^ܢ67r o 4n}yΜ!^]  {b啳LggIz˾du JSڭ3  ՖP.~J:~oߗK!IBDR!k71p0(詗q (ޝMٺгB]+Ⱨ:+VQPᬃ>yT{G,]uɧ.Ieqp|4(jJeo֯fn0Jְpeʸ/]C0½BQEY]Ã;i*PZct.RDYZVMԸQjW,<(I?Fgdgˑ_ϝK,-eIJc`2m]^%`Pw2cRwiK@ƌ30-1yv߽Iøg[; 96y`SnɗGI Nax ӻe9=^5|mȃMy@.hC7nܘt… ҎaC^8eΝ`zi39:jm2_M[솇CZW<0Fp 1IFBYE;^11>S-ߣ)1RrSKn{Ў[Dz?}b;2N;P2Tx0rm"%ٲrE"`z?;-Ze dR9/CJUNz'zAp\F'aᡐQ7֭FtsgNɟN6`Y12wKDu#t'l<ڵ+ƜFs`==d4˘g]Z5.B@//5˔Xɒz, u)N,xF]H%!x]6® j'8 0M.;ꄒZ$3[ɕ+ۅ9T牉qvu ]"7nرc-$vѢE0i>ƅ q|OVe'Y*UcOO<019r ?,g\ŗB@nYPz{~+Z*-/v$1U1[ (xq7T<B0[J[լ⮙LTIIŎ"2B弴êIcf(7ir-ޛw4HbB]0aEDuZaR :?qfdR|8oxH;jFvmTk%'e){>hRt%U1ZᡲQ/Fg7{5gRm,d`>;w)ԚN&=r1`*8hѢK4L#F?KcqY~ ;99Nڞ~~޺uO7#6VӳΧO*u[PyEsYKvzP!E<3RXF.[jO<fc""(Kc<-K e$L'hJ)!g"陜B?۪1< :ۙM|~ŋ:jd^$Q_!le\I9B<UDu~)!]tU^~_۶M3miӆF? yes:~b~AA?CL.](mڔkPدH~^Op1Re">`EQ5ӏOY/JqN޴z-[ #6 Bn$e@$I4EIv6r#?R$ $1{ G $uhst~Êu4EZ|$z4ō{<} -L#̗W_kg`)PHRETA%r&v=_ț7m4v cx c8vgw痕I6Ą{R1RoGCUmX!b9vlTjG\P\3(%lxT0zP[DL  Ƴ J\\roT̘1-ZLr2B:.sq/0!20$9OyP!wO26lĴB)fw>)^maܹ0&v1gؖPkt=ʮV#̏)\h!+)@Pޢ@w]7.fy *V f:>N}~CB>vhɓGӡV =;h1i9~UN1%-k%FP8!ɓp<,UBs&[Wc- hI=SU-#}`- K}jS"49Uk{jMKH!RWL^-Y}4=tP_JP۶\2gbb~ʕ'O^ Z#ɞRu3ڤ50_4ȔξhE4]:hKz8Yh brw`5=_2S$em.%Qꄗ1'$٧a$ˊ&&Zl|&ݻk);}zvlƏ|Jؘ|uG(ͅ)P}>P6ET{d 3Rklqd]<-(8R"S1Gڵ)gs°1S[?Cmz|"*$Q}7f@zj3 }2,vںukrϚ9UȼǮɘrG_6+\![zC>a3[`pgq(6jf$AAt.ͫ;R^R8A=SKğB$hcV]a؆ y s3g&7WWS.d(,R9/^׮QIh̊< bh*99 q>|%U#Iu""\?t朩\ V̙fIn i ۚi'[;\OMCe#YIb5#60 DT{S@¹ )ͭڳ'S>cR ư saK E!ylZt=@EV =֩c}_zE/5Ο~ʢkB_fdRU 'оp]i=Μ*jM]7n,YDŽ[Ku >z(biɝ[~T( l߾}ƌ?[A8!2mm2ůz\AT^pjz*…= KXqasAJV{$EEx'h>*^}QAϭ"~ƠuBӰZyWwAh6e!,!VGY\kZ;w 0a;6BVJLu2d̀3zPRmr_bE"yfqeɒxw&Sk_GY b3KQ.lmVt -u_^=wz5fӔMy8Dɯ*r7NnP£ Z˖-_xAƮxN>]%-CڵkL,L'H_5doy8`C|%hԽ|yE.o2m=LguY%Psg /\X5$|Hי-^,pu*,,9AB8 T~HM:M]JtE^'C IY۳GٸFGGVvgu'mfnޤy0A̬Bs\"n 'EɞJ$y@l $ Faضb42<r\L :`ضڴ)&utK? B4,HTW"UwƠmV6oذ ӧ_J:Kζ}dFZN_ ,, N_E?0Mdmd6vxLt;=@zz%@eiWCkiѣ 2bNiQ2>… % ]_ܜȅLgml?o8X2@<d~z SJ{#j4l6-W< t֬YM41ZF.,Jpk&…#0A~1C,Y\S3b<[MѣRD2gCxCu]u/Gʌ;N\l+SYԒ$)柙LKn.C'tŊ\j۷̓g2πfKa>Vleww6uV(%q20g>_x؃ _$,WfbY+U*m>q•+U;P@dcM<.AӠډ;ns3[A,DpR v=}ƴ)S&pNQ{H!I~f$VKF2e1eY,趥=Z?g#Gηnc- ! ҅_ba K_\%}F71b ,rXx}ҍKävfH1L!zN\0xY,#1I'W #GIjCJ?\X&!}xc h0$/]xn&OjlS/Ư$0?V@Yuq(= Y:5DG-m۩S%Dֽϙ`N?&=u0UV6LXs .u2NvZ:u>}=ь~ ?Ƥޛ,^(LIKST}ֵS'ʁ9sҰEί)Z9_*Ufg 2kVO//⓬}//W֪U/R y@ejuL&Ye;)<2j`ɠD7"U)$Mck[ԞuJdq%I)xbLZb k}'$:t޽*78a*т<ħG,P3~,K(妞Nz ڈ ΂ΟH|s$l)Ɓ)[u9PY &%lC?j2) mӦh9+nU%6*TC[._rs{H/?;P6|?p6lR7o\ոvWpTIܒxaÞ{0= h1w̶sGw2E1{ͥA]bSmFC1ӧϞ .GF˄x[ͰErWv>~:{VR {C gÎ:{]v)NN +3ϯjEԁ pHF=z˅K \T.ggx!NPLOib0Ok:cڵ+!ߟ;wn„ E 2ydʅ6L뿓 g1aҊv & IWR3gE3fkZK/!Foۛh^6>粿Qصu&aHs66&Ra340*,ٕv찃KI|¨ȊQCoK$ ;$iC3J)b@m0b:T5jdڸ4Ւ,2w29Qm\$3lܵF /f(HB1zxR.s˿el|s`ց3ee:e|U?8,mG1&utT3,;E )q@q~)K#٨oM|<\^y S͚E EQR6+[GjfkXQEaa 829 \u #8K. H*U [fêMZjQ-&-V$:#hovjHT/V/>3yŔ5X'yf%hkO//>q2e0M/O<jA_:ty%x9V5' RTfT_)t%C4lVLZK`K)6...[loc{s+w Cji==q{lذȃN0 C=wNQO{e[{(!2i,>Uk8b!ߞ)&.!7S<.)9[0+`Xy6RT)Ie4A8Ơ;DZ˓ԟw:wu^[B6 VKkޱk׆rmZ,&&d˝{l*ї^PxbskTd_)mEv/SFS?S73uLF )(H ~_J(rnj Y2<#a[޼ya [l2o{ 6R 5j ҅^Q`a i٦ޮ$ ݻ0#TW&S]YE_&Yees{Nҥx4Y E1矄,b*_Mu@-%?ߞW(1G[?XJ 3;zG{ĥp@DK}2۳ }{sn2kEcZLtzRsf 6!}b^aآgbw$OgYxc8B`m" (~F5k+ܻ#cl.ưf{|& ߣgvxiDZ&;\jl#&.0I#GTD#.ޚ5l C.S9/dH o)Yls_5cU?(q}&FGk F٬Ɩُѣ?Ңsbl0G^kY ;k8|d]6"P홊>ޥ)YPPμ g>˖0/>Cv`9ƗZݯٝPkKĴ{+~+Xԯbz OHCʟGu f/JӸNKzX< p`&y!Yyc B`VwDQF/#Ӧ[a lKD90yٲ$UNhIop7P%7h./XY2Z-e>QGL .!N& %<eO)I@dz `%%Y@\ӵ~}YDěߟ6_դ`xYe_W{"1%q0!3,AԀM_H^[&f|K wz*k؟<˂H,%(`a+IH =YVrzNr}T ݠny/[P\(;kK纵h>x #,L7iB};v2;6kv2)!\V^MMܵ 5{&-a~w~$ 2b"[!2!_>chN.Y3;˔ժ³`i 'WlYQb9Yc"X{M풮KPVJ.QBSh˓=WKN1%qň|ir}RBR˪FK-K@!I-xaK(bgWS3kVjYcJӐ!PE0xƿ3IS՚LIX)2m;3!j v˥4hd=hS d|ش'[tcTKn>Z5uu3fa5kKN@xAZpP譻O߻S=k~r_u8WmO:= ?d;V^\jt N͕aky]QihSxVC~O`$Ϫ4߶]6W;Om?jT7|ZGc*,CyySfOc6c`0T-O ^U9p`N߲sC]0LcHme.,SUeJY̿HV|]̏Y:e9N ] 8jf,\E>3d{ ިl;"<]!_qG4Qkz}@|~׭ݍ'm$R^(frK%J{k%7GoLZgMT+H"Yr ]2FO޸6T~Vt|z8]a $NDN8SHɻ"_:l|8I-M|.͈3Z1IF"bչ l [*~`١9A"'*qhT(!)ZsɄz-j)vJ_g}3V0QȒ 2^VscƲR,[ ^3FV)R4" Z\Wn]+<͇bsOSxML5x)T572 v7+Ξ߻`O/ydS&溋A=Z,?ILDŽٍO셀2xO)B.tM`UOԔմ[jvZDno b&!m5{cx' veElƆo+Ɋ'/\ 3ba(mAbQI:+`؄l8"{fmTR hK~D?i_Y F}HVK">B ӊ!ԥU+:r6.`!7 YnҭF HIMlGbK$DYEԀs԰@*,?z.|cKPLJt  f oī!}*VLh=-|kjޤkt Q5*K+D V@`IA@LҘꫯ$y^H5h1̀5-'jժݻW`:'JIš׳?/wojWw-vΒBB4Іґ6"";"1ݿCi8+s~(n-zW4aiKga{0 QN:.'^VDw=AoR&^Oߕ&hpdOXڳzu̓'=fwS+jbKzBۍ.-ʫ*↷p|ve,=]!*_Cܽpx0^-ȟo,Vc"[!P3ixVwuJE܋UuE1.EÚ.7a֬hQ+K˔l$iYvP hd` Rv,-"&9rFI-[b'wkOXj@K-ZE0 ZR.B2oS(̛Kz:8;vlҤIl3/M{tg92S+VH.FeqQB)˴lpc XђLeŤ{ nnq%\sƁbps5._rE(suKjs{q)r;fP>daN ᒫ_vgqiSNfI۴'y(uK0fL3P1: D{*/n .e2$ü|>@|XaT6JT͚I8_S,2.Fˊi `c 8۹s퓓ZQV@¡ P4c@Хm41nnί RV.o,wA^Q54a!I)H.?Z xZ˚:c=2#qOyq(-~c&Vos~:0,$ggT~3V-BӾlWH)noo-ZgUK_6 G$mjYh `k@s׬╤ͨxqr>S;Rh7 2gmZ˧ŷ+l!xD ]MZ$DOSQt (U+2dLV).NL6iO11?bHNE)/͖ܖΦ̙PQd ccdQ3bZ{'!F, *?喨}{/ªUo`p4 }>@B=ғ+"=R4pCB7T`74-EH!.%eL༜\H77ʥ3u l#s*pIRZFP8r˕c *c%.i֯ >!:oGAh̟_|AmĈÇة p`E<^G۞Z`υS{Tq*.\EY\/^>1AA lKbZߺ Kl&%,,li ShRzOc]# 52@-ٰnRtf$HLhzǎ]Vu>Sޕjr1Wzw5-/qu~(Iw..Dұcӭ|шB㦊K=G|Zd֨q ?YfK+2'WQhAehw$@mZJ  ]ЊoXOJrr)Ĩ_3aʀ|}%aWD~!@jբƃ+g"SmF+0ز֌fn׸qZΣI;{} "TYTnnyvz3AwaY.WA8ݸ8tpv1*W?դЖ-vq&'/O A)+yx)q0d X6haI2*QY--W8*uXrHf  Fk;,8ÆN'-!/*\ƚh4RPA(Hc~1ݻ0䩽ow^"(Ɲdr4V;;Ja*Tp4I=|sNҥK;Ef6Cjժqt\(d[I,noo/~\Ɩ-zRF1 R' ?NNlL 5ߔF ^F e:`28gRj n?RHʢ@zJv7Ca,Ix$ 2D"g~M!`aNJ`f͊u+…NU:^Ez%â lߋޔb;gϞ(V*)"Æe'RМ#%-'nX)H.5y Z Vɴz(,FnD"gR - 9VZ ,%3ȥpBoFqㆡP!&69sݻ5QPKSSҡCZ):uJa sڵkdڵo*add\xhO9%֬|5*w\$I+m)R.|օ ㏶6Lڒ%(i9 nM"n?IC|1q*[=-S_Tw_>%M':}/Gy;2eu&;$u].ʠaɪju+ R˒^u4R#;?̿};ȑ(j^`i: npajj f = fe|2}ˠm,mU8G-UM]3g6b˖2pu\E?$I=O N|0VtVs"jtkiIdN$T I>1>_ [^鴶} BԝJ?(r/ (`?*֭ޔ$M풘4mկCQj]f $I8ի'On-_էHog;q֚U9((ˉFxrȤw>ӎT_@BC{xbKCg"V漌˩y BI6plW=F mOF T,8<ʰz35(,R [B1tIDAT0IR8HARY1ofaZ[kN[2-;J򈥱 \I5ۯ$4@sV- %pܯWvl\ZX2~i:C8J\DU*٫WBLR*e#֭ *tnhIaˀ"7M~(=AS[2@ 4&W߉Յk!4Vŵ0!ya5X|׮]7,[v1sۓGm ej)O [4aVO~:+NclACQjq\&d. .#`[R>rn R#{uY6DO WÈ&˴o+ P Oɰm]Sbǿnዘl)H͘lԯ! Ӣ$0IR3k$ 7H+gO `o>.;iG0 0HG$;[8K& tr疩Ն9׮3FF]©j֭k(V̹[79g%?|`ZoÖֳ2h3:<$,@lX^*VJɀNJKβBoYׯ_O-l̙͚޸~cTϚxyZ~E?SvoSՁ19L9H-7t^-N4H'Y V2j}>H|5sQeYq+,Ƹqq H @5'nŋMjDE{eV*f~ (ϟފà[K/˻ݐaؒ";w~s#RP-I܋`1wj[`ڐ H`e2;⿙3R/Z^('VkC(Qyℾzu]o3\5󧟒 6G!4vlx3<?3`ՒcH3Zezñ KLeمď%l9*%K>kb6,xA7n\smmLm BH~fEqtNsi2TQ'.L wnD9mIe"{r7Fꮄ7⺭]+FEɗ//UJܸQQѠv{m`nǜ-K-8#M5ȣp5+I4<4C^bac1?ig +d&h sg[ mPDC޼X2~Ufk۴!A-n8e~Qpr"Y{/gZ׆ QJlW p H&C9) V!'ǐX-͊\iH\rհs;!lلO/a4&LӍ'Ee] qQWCö,]՘F:l1id={`ba'{Oj$mjD75T`acIWsG{kGuMKf`[Y44hじeI [50tK8ϵgI`&Ο ȆnP ~ *UaaQ8`̛#;"j.-ZDۇF:rK@cqnu\٥Me+$7o޴a6 MCvǏO!PYnyY@fDJWg#L>>zdYҔz|J>=/ݦML[)ȯUߺx⑄ #i탌~?[J9䢱I](9G }Ql/~)T8#n׬+=?E!9f#)o9ޘX|q$BGlFt-kx#"6BfIJ+&!o78*KC2a0}n?DeC%Ky"8#@WFA*t38}vѴoIMpCZKv H4[5$CիW!رcJ\GŇ6m8DI{$}xY,$+ $VI:v.$â$j<~QP#ɕj̴ܸ~y;JsI['(Mp{JtBs!r6o4hPMD2VCQ%2ݵ…u uj~33'MH iB)aZkNlZ|ac,kލ$ fFsI/"/'/+#)ozC˖f Q ?qDcaُZ-ij u3JlFD$v6UڥzՈamຩ%:PFsi>~X(ߋ ǪIbHLv5bb0lv6߿?|ⷘ mw\wm";ORަM]=9xdoxrzXJUIfIϺڵӧfOe3A!"lݳFґ:qzal>Ltu?F`Ћ*_ \98vEňȲUsZk ֭vX'0H\Ѭ_YrkFS}ЂEĤG8^og|-kݢUC=0Up_'"f1_>LY32tǏ1aT]@j Y"ل6s(`+vR}2g V-K6mB3gL:ղS(h{ oj1XrCj4gI qwBv|~_r8:gj3wE{2ӧP_:0iW[+˿ϻ{x EU[ @EF-ԩ؉;3e2ܧ3Ӽ*a*(Яұ46 {v$%aİ%|0?M _ cv e`#j&T:_uM%Kp Eb΍7N5Fo_d_~6ZTH.R\T>{68]Z)Pl{2Z/g[5xlҤɌ3ܰ6l斂F,K.R(c l*gzvLUd0s&KR{/R8eJ7Og/X>$Cچ +{1ʵ%\ [.t));`2m|ٵS/bcm,Ijz%o}nʹP>rz@zYzNЯ4I2&hX@BE ,Q[ҸSVm``1]s3Sjw'a||2~{Զm 䈡լ)j9KN$DW~&P72Dؠ%]WVGX/նmh7aAV=&|u|*-3Mm0s{qOmPQ_ LF${S-꛲& iҤa h܈Ƨ8BHup[ӦőLF1{ڹD{ &mjQt״j%(ĺYD{aMcå*7c-n1~_ܮS'nEaHս{'t2bF7"%Qq7K7yΝj#_ͮQW7YjZ=ly*u({WK4k'| EK ,(koR "Ꞧ0*]hgxcN{0ւ}xru5L6sķuclYJejҕ45[dS rݻǂ6WD ];-9sb)vA+W"$JO>H2 j[>kNOGr;w"C]֮7,J \C FolT" :d wj$# n$u*ap*4p[&! P Eԣ aǩ Gm8gƂEzlL( !=X:bv(Fc Fl½3:|3} ]':sEwl(qp#\@%hD@>Dْ[IvW\9X6# 국%|5LL >R`Qm3\" `SdY&(HnIYbl ĦZjb1`22V)zZE+Hq_OоkJQ+D@'N ˲f͚`w2{=z4˱=% M%deP)YR?~_u6=NY2SxkÆ9{vqZ Ѿ^x2H2Y (@Oپ={L5@'qclD׊5jb v7nEb,;& H`K9 %n;wZKݻ7S~V%5f߶j$V<yn AcİݐCZn]TY!xƍ6`&OOq>*@|2e:oW~A^ GU=ZV|њ'{S }!S]jCرc-]*Sf 4(X#>Nɰ'$UYJ|mv#&go<̂{3 u~k 0%j1M&A׿%f,&( AU{ $\&ǎϙ"#_6=*8^<rL3=z奚oc80͚p}?"RJ/[Ftg ƍ]۰cO0, CpDoqlm*R4r`!Jq>˃  $ZC\7@%c4-IW 8"`֬FOnpϞ(K-ٌ!X/}ݺ*td4tx1lۤ52eb5p3 L+֖lKm(%q 3.UfOs:a+& 1 Yy(dG |~ճp%u)%rT!G^gu[2|Q X5RClrƍ#F?Ky~ ;+LWjMI'N(9ofV_ΰtϟ_=͋rp fʰ?ZbɁs0sulA5$[FN$Ixwުh\jOwJ4P1-Fv۲3 d 2:0V1ݐ?|8i+ro-I-~ݍ1dgY5 \+|K2-Ā͛c0 %w!n:SJ2dzc$'L. vQW aCﶥ|o6_/F77`I~'HQ84VfSg)Xn·4i0\\\0l)Jc$9ƪI)A$qWԟFwRoT=vMή|pP,JUNjl#/|2@*22\ngˬpxcB2&yPG}_X!Dn\8.Ie+-{!XrH3KKd2; Gmώߊ]zpd2y)ekެنU.?|g2 V f2<8Ip,d"6v`»*;Cڵk0Glpf?sfȑ;=$M5Ӳ!ե &m[U+%"?>(4``94|>QCA9$us(EB!8PpdЃ y"/vhX=2˜o۝:9l;Bcaѧ|g8ґ7Yavzg7I3ix^a/dX;c)A#+ |& î;ѲbiMR&=^ŬZ >V^pvYCtS,qO=w.Kw*T',(SȻ@s|- `,!!#,ȰS0܌Ξ}M|8hyI*4o~ϫZ<5ZEGU ɯRqO!? 4p&'j/paxWZ:z>U+\[@ o)$39viW㻀9C/W.]pNMow^N&\W(EmZj>p`sFLОa|#''\Ze MKw;{ro(;wG법 .Z-7ӧuxhѢCB ej[Q#Oq魯 肰<{Y[x>d*Um;lY56ԢkOLחGQ9].jCMv9eJ)O.J zf2"xpKqנyNZFUdWmƏwڼYâHȻvEzK=ZRF }-"x ؏!j$f,%'y&T5lY}uv0Gψa8=ƇZ&M_ 1I\b ̈́dJWgȅù[pֶmCS=dϞ}ٲeeʔyf ê1oH.W9ps[9xpVDMtֈ>5yXad5> BNW.qoSd5<(I!O.g0w5o>Monq2]3^jBp,W, YVAdG T MNM Yށi};j|L&NF0),*$T+*NMfO+ds xr\PS(czĴmȍ豚T*s''4dj+ֻj"Հu7/4 RE }[76Wy!jty FV 4RX:c8ۧacTBΜƍʊ1Μ ܁ Pt G2HR`?bul9h* VbaC<KabV63=,\bĚ0F:\%JB 9szl_{7yxRtx r"߇94*] Qk)Q8㜅I*'^+_{Eْ^UHP E@DED+ UDT{E"H(R@nvdp]evv7pɝ;;s=3zbpkQJ`A:4o$s4.O,OH )EFf+\XU7=%LzS+[*UH67J0@mٟ*-$f:8G8ԪUl2"߻)~hxBv0a"+VP~[Cg4ÍPL<*EׯPr?lR#+!aa֨.2\@ѓF D6uIK0 OffUXYJJ\ӦQ<ʮÇeS ڤIi=# ބnO-[ѯ(6Lh\|3=Zk[5 2(P!=oI;$Сr#N=]jaoE3/.Mo6ʜRD~/SctR 3T|ؓו:헣%g-}N_):&gNJ'e-tW L:%rt#̆ 1'JԭЊʦ% y]/Ӿߛ>OT 5퓒 Z %5hArR+.qEIh$'$w5pn٤F*/?|3}nwݩ^\\c/ȋX44.jsB[*]16>itasvXq?Of#d ~1#tSLqE2i.,X r \J"$H @2(L20ėW0fTGMcbV\@A U>f!.S10:"{; K1p\ti7)q&}ͧ":bY^DU遙g-gԊC -6W۟ %Y3,R=K5K_;+wR6ѭk 4>CU((!f xһCglۖʥ+21:n\Vڱ[GV E s7.L)LMrdFBus\r MblJŊ+\0̜|rX~w^aT9WRd3+?MM9jEL0{VU-{+Ǖ/ V6ݪ{"4ʈ2SYu뚰H:uPL4; U2ƍٳ"R(ڎR`s0Yѝdm7H1RjHW{<|{|٠6\G +羒%~--U^fUITnEPO?4~y PQk֬QyDWѣGUQ7M0>Ҽ]]_!u HҙO->ɋbt[Q7OZC ~IŴZae+M'AQӺ l3Ro~?x-2j.i톷P2~}(YAg y, T<{pf+o O}bX04dvލ7a{p@lԅzK 3̧ :K+bW!VrQ_=ӡ綞_#_.$>!V $74Kn[d\۷4hNĉ?X^:ˬ٥-GWÇxT;_~zEtƀ *>0zFDcҊG'l+jXCu \FEuq ftiҮQ ZJ""<*y پuZo_l3R Bf(G4\xAQ,ҥ3~Ń_Wf0Ѱ̀j'G |<|Q`P`1p)iY|9([F#Hi׮/{)} \^i\B}jd'G43GZ#i̘4nuXoʄ,!r%ZsEٲ,H†ع3Bz VdWˤ:󨩔 ^)2bDLJ$v_ Isņc<))ʕkEo8vL%]. {D`rɵo+~P2AOL02($52*6NhHi7MS[(= Fmͺuxe$Q5i^rvYc~N<__+)@ت3g\[KbmH-ihڴ)k MNGTO1a s:- NҒWYtꮮ|~s7̵X>W9]-#=Uha37ȁl6HV.5S9~&ޠС {h>^niݔtXlߛ聡2Gق=?>73roxժxe @ڄG,^ Y6 89Hmh5cp/@=xp8mmH/(y|;%Ktͅ eNFF;*^LVYO\oɺ.aff1C-A֒!HOhTcEk2&5ߓY/*D۷/Yj@xG({LJE`# Ԁ4R'۩}T` TB,Sh53b~u4Q2O2M %N8US@ҏ3b/LII9vdpU0j7{r"Ɠ)ז>yRwtҵg[xIzh LFΗԕʻDMW^+h'AR{ș33U%Pc)2ZF.l-+WzюcF)0鿆1C`_mnZ|,IstAwLK7S2t鉦h%O+=}wƃ2T뺱h喋5:ܴ91lUXqɒ%MM8_cۊV^dR_RA~ܮ<"*^#`.$:r%?ø]Rn!u &hE)Y[ȊcZ^f\5**Ws=eb_'١aQ Nf}R%gOrАݪY{rҠxWw͂n$!t+3Y!Td0Kx ]E1Ν4NZҔ Aʆ0Wʩ$UKa?r=1+<]*UK sxQĥX;!rlf 5tLjǪ_Yw @( ȗZ۳ۃ^esB|hi<$u@?~j9Dcy/M {POb k_5~°m *?yOYʽ!^]dJ A3'ڒ|FL1&Mŋ"ףs%* k߃EDnܐIuR6B6rui~b# ϭ1Az`,)X{& s"U8*pWԩ]F5o~}ڵaoe8v  ЉC>j[#P (5'{-U])a.GE 4?У=wLS/ׯ\I޺5%6 Ϛ7FQ-Z/=Ktϟ |^`ĉ w$:$ nݚ~X(CC%e寬aÆ)gƨ7w}8U2Z7!d]R}+J'?P6Ӭ J)J+Kc(a[1pŮ d%|=ڐCw,۾%sJ:c]rI6" 6TmwSp`пpAց, '?(cpߠyTRt~Z5fԨ0Btܓ$ğR,(r<ģ;xǡ-[FzS1OX.KrVr X0_o~y (J%<ԾU+2Υ{=3°z]ЕK??fcTN־MH~4ܨax-F##U|2^yC )7/\H4hڵ0#3؋׹ jGtFE*p/ .Z9vT`T@`3$Tٶ2z5&,:R"LYӉwpo(0ԩ)uS|}ZQ2?  TUIQl( '/տE[9+#]LY,m}6f [zw_<"mtr@Ŋpp`Uoab0ΟOdE\ccC_8Or3Ba8p2F Ĕ. !hWT ߣ} 7@O';sHXқIfjMPKNB\ V|A8ؙU%x•`ߚ._ZJqVgB&Pt嫝={.Vg*#* \%5B]׃w6vِ..EQ i]v/Ǒ}'l^'%U]{tD^'NLyh+JA#6lX[dU őqqx, TRe7NYCխfU>ʋR Z3PyL GQM ]c6~ 9R r`WpCȔW-# P{I`F"+6}txf SjY=/]3EW($ᗍ%#{ }OBFV;٠xwW_a\ǁotn/+5v 8# VCicvMY X#"AӇ\W^Pj0oi.wg㏦Hi,D=_aXhBap.ʘ7 NHP(5_wct VZuҤISN9ۻd/^s\{x"<db5jdfbs *M=p+` ojJIS: E_PR@ѳwP2"켇='n@I.\~WD >m%^ȬY7/W*ȤigdG<`Ű3N> "-mB Ųt IY 8 5j j]R=|)Sp?}󍌰E駡_D=̦Mr ƒA͛V}aɅ^CBF2Uy &(SZ- EGD8<~ ٗKmy| Mآo%8Ggp2>qkXDR/\%5B/Zzmxx06*\Xl.g,Xi^ݾ?n1bՐ1)U\n ]oyJLnoCI>#B:ޜ;^ D܍U@^)Ox$(A'Nl:thV "Ynmu!.lR&A 6. VjX+ϝҰa4Ye 0!>EblFXk\Mb6|ž S@)6a\g4rJ"j6[̾߃Ŗxsm$;wŸRouO~I0d4#8@Pi|.yhrBrswzsne`E\~p =ZqNK C{?=y_řGJMiтF;a~~ᄵ[ mz8G5n dаҸdjĦʕ{GbQrjmVS(yټ8 YcN1>V끴ׯ/#^ r["Ɂ}3֮mQDfjM֑# z(yNKѢC %=z`p/]ĊlJW$ =utt|͟:RC:*-9#8R묙z0s]. .TǠyܕ kMÌExMnV}- k~^.x }PN:q&=믿ο+WΝ{"CA9A50yعW[rI6'b :E6YTǎSk.$l5:9[ )75^7eT*r,J1A ،!f0HRX fժ3`?p`:A~G})`6;Q*!5oذpjŋåNHHuqqr{k+;N g,ed-k#'Cj4(J1MUzdT#(wjm il%MG黧oGJJCPӥmHѭ^-]'ꦭ0MZPwݝvGOh8Ce%5"=;d&.n\@ƿ2$)JKS_*^J\N bm $RCsl644` /}P3yz/S&ΝW}BV.BkFgǶ(6|J(cJ|-:}zvM5>gBO}<|9wS'9,`(Fp9w/\G1͡>e9CV_{-|9/3X=|p*U_"<?w*5~ƌqqC!,Y4H'cG)ިMвm%ջyҠ7Ed`1>*_f{<W^Ev/a?54<[DEV \clNy4e-fW_ݤ;u\3Dtq y7BI0x׬\( ~`M6-XC"$~٣B@\\mڴ!~U س'g5_;A;w1_~/!!d"fLL&M]7::WqC bzq,drPkIϺ0P?먥J>jא׬4#;f}QRYЖ>zTdNpKN{zw \Q#;+8p`܃*j,Y}bscޯ_O;.1pQ W_<{X~x_:kJ.Lgު4ndy583Й9sppTzYYfJXQit/Ɗ$>Е۷yqͰI >^9Ȉ.Ν 7`S˧Zz!RC4L41G? q qvm*NWF3h+1%3Čݻ/;h6myX.uҤz3\!J{ (R~~YCi1s>]tX8?otmi1\dͿ~9׽jHl8G8Sh;ېI\i8*g/EJW$!%8G8ޭŋ1+[M|t\_q\E]@`d8J*5=DGIҖU\TCՇm\F$Oq.TK ,d{ˆrO;v(0U]yZt$6 }xb@T|9 3]tݺ(0NVufa]?rrf^2^xA|C_d֭HmR/1hA|=!{6nZye0*)ȍ՛^զeo;?2 :c,h+&2(Ñ˅8ijA퍢% c 9˜LТ8 "A::m Pv2,.\8bnk¹\(O 8M$ұ@샖.h;v,dK[ y8wPiD z8?gv krWDZ& [lhSUUK~}tZǩڠyJ MmV+39xɻ(lWPQ.H%-,!F,BNF-g}R _ 5G\ә;vgD"# JueN+JR{JwQ.wȅ9KC$W\X`,<mV6Y*f$rf>&C:txt'L| $=}(EO??.zQ[8Pj,'8{Q!kDAJ@3mhx) Cbj,t' EkIKH~P՟%CMg9Q)=.LdV?W. # q$9r$/\| $y 9m8K#_P"}MP1 jAv%el r ܱخ]n[Pc``_/l- %DfwHZz8L(#9Eޘ+.##QﺫWY~0\x;p .9`اXtlv1Fj,6bMct/Un hz]ź@'9Qk&wXuL=o7֪@J ~ nQ#ĉi"$ܒ -+` Ll- q5HFy5˩Ֆ/_c&iVyj}@%XСs(5 ڊФ$U@JwCyl^4_qƒ&Σ q.]X4r=kS4@"Uص|GݳgŒDeɷSv Q/6|8ŠHH֦pyڨQru󲈶N**_&,-:o,崘'HHPTJزe{@JDq #V[a! +w5#^wr{ۿh/$V%@ ͳ?v2$`ș5R4^~e*ǠA8p1rxpgBջ-Jxnj-=ntbj?#Z<#.uQF tO/߿L6n\$I|݅uCBRܾcXOrWlyĿwޭ[H/4-q dhy`^!fʠ*V̲nqLlLW[0xYј]lΝϡ TZVլ(nI\F>nݺ9`s|JwߑU:ZDPi7ε4nlZoU, )5C5Jzo4BeV!nϞePQ<^ 9wek.t44=?ߏBȴ#vyc\_xet5JYӧ$<׿b KEL}*۠f@Eo#nə*9#Y[ߦM7hcjpwkɆ2n8γD>p=ApR46r+w駟@ߩbb|*Æ /#LոkiYZ%5-e|;fR[ȳ.eV?}{:q5j*GطMY/J6jT /_5*l|kcQ\jI?O@O!hȃ4f.7?4;D!:٪x8h%r76?:%<u)Iٜc gReh*,/􆱮x 4 tXlnbidjVi|ow估jb *%mSR]n5 KU"NpSaY0 2VSAPiȲO$+_p([ 33]K654T㓤G={'|;#QLĂ{[ŏu.23Ղ) d "A9¡Ӯ]64QH;l޴@1w5Zc1cc HĖfCbmK,Qd~F\80X/,-8YΔRJǏl?ڎ*?B. ˑAG!@%ƄS04kw¥ぢEVڵ+ ;l=NLLtxx;aPY'Ԋ󪊍vOnpGDQƆ|sϋZ]ށ$xwҥeb![CօVzRtUa*BA68Bea$U9sFώl2S~/ȠշN\ȥ do _JpYVT͛1rȴM?QRbcU%2kVӦNUGAy(8~r9Nx{nWYyNpb@Nz^wDrΟf l**>=A;H2(HUͅ^$p 믿J6Ѧ*a YpJ2<(~|(+ݢE sj30Ȝ/?>`orYm5h J=؇z}a;o b\۝&v(suxƮUV !P=hnaKΫ}-זA#3ĉmn7=/]r=>"0X9^4xYDNhA*mԨQu-.%yPfH 07@ *4D4P5 'OڥWVٲ:O:tlH0,)σ*$K&&h 舻أle6Pɬ Oj^. ;wH JPc؆3^3N>9"8 \&.5VSұQiR ʺ@da4R<ρIAK tT= 5[2O=oMVJJ72cWU{s3͂F9icXqG<n$zj SҰ<{4Wh2,6r+c@]*BӀxT*xUVCaEĢ0'5ǣEDcd6/H9Nh@2 ȷ7U'l5 o/ 9n1cj߾}eoFpŦ"t9W=RqJX Ppy#P9H%F)C]ܳgO ̙̌39f CHKO'\@wW SM?hMbi*RD f ] tR9ۮ[ʰsL>7g!Fu[H@ÇM1b[}/ *$Vk<~ t%Λ)G‡Ӯ D={+GCotwF"% k5vd>W<=j* #!Ws8n!$^vT8B+k +x{4Ɵ7IIt|><1ʁR ߱/^8ГW$AΈy8Ej1/f.Ԩ" B/2:['E , ЈQv*І8Pt4`;LbӰD a]wB\͞-/GݡI"$Az v[<,ODԴ1N9o>z>Xl ԭ;UD J4XDpӧO_/msF`} Sbi.U!QG ۔@{ IRN~:*SP\ӞCt Vw>w'rI@zDQKm#&o'8`(AQp*U"H'G`߿?/ ^C%x#ի>F~Ʌքi!*hG,[lT ؈6vG{T!*iz;BCga4yʗ#¬~:EpIDR ZHٟ  q*8P{4 R0P9 ,F4q8ձ/() pZͱg5V'fu.d˵~i+c%$w ùޕFvrsIf0|,l57kP3mkӦMPDםB0.D @Näpc:#1jԨK[T' |Xi@& @n[ๆ_^1%Fh]PJzllm5RGq Ϟ$A#dx@Ϥ3Q* "7,'%%iXPuF q#SQ)F*nߝ~L ̈'OAsGa%tZ-ҡPixtVಭ%7jf:PV{5ƚOںBP\, '2lŕApZ'H swvNAdɄ ̋*ný〽 R ~|!Wqw&q&=i=`s@KX\e̖XFP~q ~4sm"$K 6=@Hl765T BK|D - &HwP =@4בO*P1˴~\}y %~F0lf@P.;tư{wG;l%ɪä?0>JuQ#?Ј 4޾HbL ц exUw>&?$ p>.=R'&)|%h.kDmSt__\6o<{b%4p lZV)FҀ8R".9qE.YS %Q?>sWJS) fz5Ŀ dԜBp\}QI|!K@f7fzOYj @Kl(/Oe#"FJ,6x#ߠLMn(xO2Yll,2NY}J@fŲatKcJC4:*fwWjc<J—@@XĘ.[Z @-U8\ӣ؈y. c_A#c$G8Gņ#iL7#+YjbFV)LmJ ,4V믂*7hsߟ| =F^o^h5裏pgj_7 ?pâ~3/+6^3^9#Z!"/lOjrLl9V1%tĮ1<:V H *Rt'ēÀzf^MʳeT~P}/#ُ1csHȾ?a-%$M^xqŁ}F)UL*_3ϬPgI/}FROG!!bO.2I -*n#cp##1_nO/~9*w|^uꤰ7h`HؤIAV?~ٶm#Yb|&gxщĆ}vۅsbШt-j̥4852x /( ԈA)"O^11psRV5߾9[p j^EU%+]nԩ6:eJXjWX\-tj~:DvV-,$ڵK2㺔1cOګ4z'1K\BB%`U+d֬30,,\(Txg7E*oI@6FbL={f=@V@@柵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3ZMBBBBGB埵3; UIENDB`:Dd(!uuZ  C 6Aumls_conn_distb:~ӟG[ʎ9 n9~ӟG[ʎPNG  IHDR:q%x pHYsgRtIME .2 "tEXtCreation Time11-Sep-2006 17:46:50\$tEXtSoftwareMATLAB, The Mathworks, Inc.R IDATx鲣Ps+t  kU*5t}N\W (qJ\W (qJ\W (qJ\W (qJ\W (qJ\W (qJ\W (qJ\W (qJ\W (qJ\W (qJ\W (qJ\W (qJ\Wy{>v/n %iDh %A+@P ed,x[zreUTWKďeeViQx]uvi vژU]Vf=r<@u4Y!Ai V^YJ5]ݱU],_6s?mU~Ӈ}O1wχm}O-s$Uh %A+N: dEB̸2 ?lӿ~eg,]o}ue3.7]/.STWC9TW (qӿ~ (1''TCDi %A+@P edj.'TCDi %A+@P %A+@P]i"b|J5D&2`@P %A+@P eޕj&.&Z'TCDi"i %A+@P ed0эf "Јk ͧ֓p/zGu\Ln2+kE Z]vdmgj<2qr2=e}Ҁ4D3NI`Ƈ"R 8V;Ϻ]׈4: xI!+@a}XWO'+p#%R Ped0|)TWv^յnz+DyDb!@ahF Z&s?% N|@1cb@ħ\G&E+{EWC, 7`=-+u•Jg_͎y9Ӌ/\%S]ku??N @J\[DVٶO -3) qGsu ܩyTY19[? j1ʼn+@P - )J+`\1p+U%A+@P %A+@P %A+@P %Duۛ+!*$뺾 G\y (qꟷ7Ǝ@e[c:t).3Z1r]"~,. %\qGB f4;mjj,p*!Օdx1X5ġrYJ5lA+@PJySN$Ri{Y%rﻮCDMz@c̎NaR]ɒ!"5]We,7(958Օj .WvD,`4#Lr!84CuvXSyCO}ȐOW`B8(TMu2Nd}^d(|J5D8`4">=Bkdf+@P %AV+Qb|A+@P %A+@P yWaHh%R .̧1%A+@P ?td c>m * Ew-*,+@P UHQ`Cߕjt^l}L@)A+Q2~^.Z@ >m']'D @P %A+@P Ř ,qJؗN#GqvH#oW VtipT7-DvB߯/c/+Q^!ͮߤ 4\"!#ik&1_%Q (qJ\,]vcPnp kYTQf8H\I/[V V7H `%@`J\2$#q"%;V0eV{x͐ONPX/;U>E\fĜi Vnrk<ԴqW7t@p+Q>lЌr`P+g<y r*W  `J\ S&qbJ&}?<#A8STW (@\W (qJ\Wn]Ƽ+P1 7 | Gu,(WV!Ehꦢ qJ\*- )J+@{ dnwau󠲼R˺F?ax1 ^&PX+@P L,@ĕ@t\23MP`: cW c+flP\^V!_ϛ~M1VP#}u]э@VH jrVU]YgHrWHn+ABcV;/ugi rUS`jrB-VաuWeӱ–$ 8E]xe+Ka0xFM@c05WD&:d,𤭣(a@XTcp J>q}w]7;@33B)}O]I/h v)lGb0gQ Օ,"S2 DPf"WO*h V #QVb ĕj(!Zq`4WBYOh63 .qjQEb"A2['TCD[=}pi sv^!*% x̍/I,`*`]b0g+ꊪ q8GuxyWaH5 pLO\+n+8* 44%+`QW+>WN(H,Mਾ x=$T4D5qCLO\Bd%D 8G{1 p>3 }++0S+EWn%wQ`2+@ I\=韷7u_WkL,X]6q" Giy! ̮ߤ 4I,) 48H\JA(H\ ;2ս p"Xa#.G, pĢ1pNb{VݙA/FEx֧vGRS\9T[a}ju TV&@p]Ae)އZ6v?͎kuwVR'}.ҟH9 sEwAwj+Eޢ=p\rSLn L}@X аo~=f+ךou]D"x`;XlڕŇ~QoNb 𫴺2HG [S_Wu\m$A WkcѐLvQ1`1N#چUWVG^wT_\}*P*XT > ָ2 $*PdSuǕ GY[K,2J =qNX}R 7|LGwP]#?qƞ15mަU_ uӡ:R FzPrA>.Q`X ]W%uhD|V@o$j]- 7eK)Pƺ@ǕQ"@E+%(qtȏ"vfH,ОqoB"@u=G}_0]]?4Tm^]ʬ(Ќ2=T2",O :@+S|~,u]X:{TWZ!H{6x4$8O?W!/U0c|5xxFqe7N9PV}qeYS[$[d`OB\xHD3H@j9f+O8{۵8!ew f qZ._%S]xbM$2r63x8Ku9gk'Ul?a<-eunG}x@Q+8>j7 \xj7ht_]MAz_V;<|1 Ec6q" 4~lًw]I/h P/A4qeWˉ#w$! 1/+;U,s$@3-; @v?؟~ *`v?ޟ>f,Agj z$vO,;ШǪ79n+!L_^ ) @, yJ"g/VUT~8K\j-,{@) @ݖO`9FBp…Y _sz0V+!_XTN5 VoV @搳#`P [ ŻG\xظ\c/(ŏ__5"WrpC &q%/yŝHh TQZ 2q [Stzĕ )uXe`ȑ쟷7 ;=\xc-+/u-;s\*{/ +YȱU22*-TTW2+lSv>(Bu3Ζz)+%Zs-.~vgُ-}d8E\존^cL, ;1:S8h?k|൳ʗwMTWP^4@fm?m5К9XgMTWNUg˾"{ e+q]Nd4aJ5@NM8?TY&ՖfN?K*%TC,f/!sv4 qCL3@]̻R@` JLy[&Z'TCD5df #a|  -qB ]/)bxǏWk\#?Bz." eLr$+H iEJ+L2s+0|i9a|,,׹ E #q+̬&_$r`NEil+Qh3'a6~^. aNW޽J\Y$=1$+_WB:ޛ@uV抾x{טX[ qud;\@Sha ;V+$Gв,(D\`2W,ZegۧPq*MWXw[K\` v+@P @x;]`@}v+@P ;AX/WwdtUW8jd`O$v+l#@PcPnf&^]' m.4)hf ^5w.AOpDuxOSVxQ81;Kun1}!_6~[vj +3^Ν1X5G8E?{:Wn}u43c(@{tOw?Dϛ]I/h v)lLe+Y!DDX0W~b$@ETWxέ4PPqJcj ō#7D$}+QeUdN Oc0 (q*n,sdww(Nc05䊱ɵ|+?TEud"Y Ai V yaPcn'TCDzwUn2Χ1k^o(TAc0޴:Nlv`XxꟖw.o #D߯O[~iYǘ3 @:u˞e;;oqEw ƒC" 0M$м:0J>q" м&tBa|A+Z([ QZ@,өW۳;@ө!Y K_y?p\䚗a:Th2ukc򹢻@u7-3ɴrlm oW0^vIf9ъLY 7О|ӥ7D J\Mt┇7`kcx`lMe RgSݹ}e6o#lm վF+ 4W] i,em^.z\t>%}wrEw`0ݪ%H\w:eDUM?׈+Q/89y9-ߤ 4W%f]bت%H\ y< ,9݀_nc?(sEw %A+@P Y<ޘi0N95\ҳ/[?>tۆ˭m׸/0Z {|Vx$𮭈uQBJ@c0J\^ Ufp|{/!‹x U_5yNu9B nwwppq? acξKU]- 4W*Q@ح- DBL$^B k6tYCļ+&Uϻ:Jzޕ)YZ} P;Wth %A+@P %A+@P %A+@P ?oo_߯^]ȝ]wWNa׭[;9u랙Z] Y࠮w^ASSrwVm f. _%0newVmx9UK%A)Hաw*{ h_mwaIn qwR-Xv (]5T@Βq. J\ /};6ċ*zW%WgLÍyj޹SWӭݽnfoih?W-j (qJ\W (qJ\W (qJ\W (qJ\Wy{u]7-Pĕu]7(i #qJc++0W7 4{A CdA+YsVX>#TW2dZNP].?U7M$R]l(gT/\텇|bZ]٭~|~Oݷ1M8Z{p+MȯZ Ppߙj?}D?WH?k˵|{V_a_ekiDj\nV-_lfϲ|-4[t%|F6l-uJ OW9^4mB[KFe"AJed,IZ<Us}+upd l)CöGJ!mo J\&[?&~:jN~^cW;|!ͼoWnzt\`jx{r%Nu'K+]/9{3gjwC{m厞咳MއʷٟSuQ_jxzq;`wY JcL[ɟu~q?yS#X%Tov"s^wAv9+P%W-d @\R_co B[髶J\>ewߤ**; =N[TW*[`oJ|ؿ~S_TW0ʰ:6N~wҤNoXb?Wmy@u%1\.WNjן{=I|ދWޗr@Wג]_nw f@E|o1v\m8/1ӿ~ ;o>Io?N&̑Av7dM} (qJ\W (q%Wv7dMe i"+}u8.N&o߿~FuCّv7dM IuJ\W (qJ\W (qJ\W (qJ\Wy{2$;kq,O[j;/knLߊ֖C+@[fZᳶ8+D׉+@|A*ˡvP] 鼢{)TWk*E (h8-g=22$v,2TΩ-E–9~l59g3P[ykyr\ژ;6lFu|j{ [?2MJϳ`[|>$}!I9%WT9sPt7_-)"-ȪҍC#(ro?]EhvH\ە:ӕ*6-r-;Vl3~y.=gϲK-*%iw%[̇|ȑ0c$Mk)_rVS{>#Oq3q0gGo%^|uP‡iK.PwHgSn ;\|>phknQa߱ EV-⭳D&@"\ݲt%nt)N߆#?ȁLYJwXዜ\ع[kjj,Otx e.pBX ->o𺏼[=vw X]DrvUOuή#60fޟG/p7qӦs~._PM*~j&h+Rd;YkpӐ7=ȁ{٧sPUww 1@[ZnpKb޽NZݳxEV[[g L\sN쯘<ތA-_l;[xr6BR&YI3ODιS+mwvVg'H;.lFXE^K#0g^؀Rr=&iYj\kwӏs| Gu>j;.|ߴ%[F2 ,;&a,1+ok D|-w_&{PR swǨM}&l-WЪ},QK!Тar6xPJbK¾[_K0g90JYmS Iq:[bDC2vuW&lvB [yi+c6~m.>yQ<#^θ<(x p%O{wˋ$w9~9 ]Q3"֋>Gr"Ai %A+@P %A+@P %A+@P %A+@P .顢IENDB`SummaryInformation(DocumentSummaryInformation8PCompObj!q Normal.dotBiomedical Informatics3Microsoft Office Word@jn-@l9=P@& |՜.+,0  hp  Columbia UniversityJ 'Semantic and Machine Learning Methods Title  FMicrosoft Office Word Document MSWordDocWord.Document.89q@0@0q@ 0q@ 0q@ 0q@ 0q@0@0v@0v@0v @ 0@0i@0@0@0@0i@00@0i@0@0@0 @00_ 05TUVi$%&'()*+Pdest  T/ !2 B $''n()+ ,B.0033J58W;s;?-?\ABCDFFGIIMNN:O;O\OeQRRuSdTTW]!^o_`aa;bMbfilllll oLr\ttyy|||;|<|=|>|?|@|A|B|C|D|`|~~b&()I݂ނ<<bESTԎeޏYn֐!Q֑בlX uRntњ#B)rԢƣ9nQh r aSHǰ KL"ABDETUWX()*+,?HIJKVWfgh{|}~жѶ#$6AVgklmp000000000000000000000000000000000000 00 0 0 0 000000000 00e0e0e0e0e0e 0e@@@ NormalCJ_HaJmH sH tH Z@Z h Heading 1$<@&5CJ KH OJQJ\^JaJ \@\ h Heading 2$<@& 56CJOJQJ\]^JaJDA@D Default Paragraph FontRi@R  Table Normal4 l4a (k(No Listj@j h Table Grid7:V0Z+@Z h Endnote Text 7$8$^ `CJOJQJaJ>*@> hEndnote ReferenceH*e@" hHTML Preformatted7 2( Px 4 #\'*.25@9CJOJQJ^JaJH2H 8I Balloon TextCJOJQJ^JaJ6 D!#$%&'0C1E1111111667i89;k<<@A*CEGIIIPXZ![\\rtuvRoye-_{0 O 6 ' )$F{^un`U !!H>AQT%&'(FGScdxyz{|}  io *-01234k gdURA@=<:987 H>AQT%&'(FGScdxyz{|}  il   !"#$t!ll|To 4 oS  B 'n()+ ,B.0033s;\ABCDFFGIIMN:O\OeQRRuSdTTW]!^o_`aa;bMbfilll oLr\ttyy||;|#9hap@0@0@0/6@0/6@0@0?@ 0?@ 0?@ 0?@ 0?@0V@0@0!@0!@0@0 /@0 /@0@0 K@0 K@0 K@0 K@0 K@0@0T@0T@0T@0@0\@0\@0\@0\@0\@0@0d@0d@0@0q@ 0q@ 0q@ 0q@ 0q@0@0v@0v@0v @ 0@0i@0@0@0@0i@00@0i@0@0@0 @00_@0@0@0@0@0@0@0@0@0@0 00! B 'n()+ ,B.0033s;\ABCDFFGIIMN:O\OeQRRuSdTTW]!^o_`aa;bMbfilll oLr\ttyy||;|p@0@0@0/6@0@0@0?@ 0?@ 0?@ 0?@ 0?@0V@0@0!@0!@0@0 /@0 /@0@0 K@0 K@0 K@0 K@0 K@0@0T@0T@0T@0@0\@0\@0\@0\@0\@0@0d@0d 0e 0e0e0e0e 0000000000 0 00' 0' 0' 0' 0'0'0000000000000W;0W;0W;0W;00B0B0B0B0B00I0I0I0I0I00;O0;O0;O0;O0;O00dT0dT00] 0] 0] 0] 0]00;b0;b0;b0;b 00l0l0l0l0l0l0\t0l0y0y0y0y0y0y0y0y0y0y0y0y0y 00000000000000000000D|00D|0000 0 0 0 0 0 0@0D|@ 0D|@ 0D|@ 0D|@ 0D|@ 0D|@ 0D|00@0y00@0y00@0y00@0y00y00@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@0@000!0@00@00@00@00@00@00@0000@00@00@00@000@00@00@00y0!07y0"05y0#03y0$01000y0'0-y0(0+0y0)0)0y0*0'y0+0%0y0,0#y0-0!y0.0y0/0y000y01000y0300y0400y0500y06000y080 00y0:0 0y0;00 0 0 0 00000  05TUVi$%&'()*+Pdest  T/ !2 B $''n()+ ,B.0033J58W;s;?-?\ABCDFFGIIMNN:O;O\OeQRRuSdTTW]!^o_`aa;bMbfilllll oLr\ttyy|||;|<|=|>|?|@|A|B|C|D|`|~~b&()I݂ނ<<bESeޏYn֐!Q֑lX uRntњ#B)rԢƣ9nQh r aSHǰ KL"ABDETUWXp00000000000000000000000000000000000000 0 00000000000  00e0e0e0e0e0e 0e 0e 0e0e0e0e  0000000000 0 00' 0' 0' 0' 0'0'0000000000000W;0W;0W;0W;00B0B0B0B0B00I0I0I0I0I00;O0;O0;O0;O0;O00dT0dT00] 0] 0] 0] 0]00b0b@0b@0b @ 00l0l0l0l0l0l07t0l0y0y0y0y0y0y0y0y0y0y0y0y0y  00|0|0|0|0|0|0| @ 0| @0  @0  @0 0  @0 0  @0 0  @0 0  @0 @05@ 05@ 05{0/1@ 05{001 {001 @ 05{011{051 6{00{061{001  @00\&{00{00{00{00{00{00{00{00{00{0 0{0 0{0 0{0 0 @0@0@0@0{00{00{00{00{00{00@0@0{00{00{00{00{00{00{00@0@0@0@0@0@0@0@0{00@0@0@0@0{00 {00 {00{00@0{00@0{00@0@0@0@0 @0{00 00!{00={00@0{00ol@0{00 ol@0{00 E@0@0@0@0@0@0{00{00@0{00@0@0@0@0{00{00@0p@0@0008s00008{00@ 00  t 5 8+;GNuWez) D ao`efhijlmnprtuwyz|~& ;Zt`tnSAgoacdgkoqsvx{}nb24 H K """#I#L#333?)?+?t@@@e\\\|6|9|/[^DpsoHHHHHH:::H_HH8~@,t~)(   ('Z) +3  s"*?n , c $X99?"`('Z)h - 3  -"`#"%  ZB . S D ."ZB /B S D " 0 s * 0"`!('J  n 1 C  1"`%   2 s * 2"`"%t&(  n 3 C  3"`  n 4 C 4"`,H   )M& 3  c"$n  c $X99?"` )M&42  X0P2   "`x(P2   "`Xn  C "` 3P  n  C "`d  ; n   C  "`<#  H2   #  HH2   # H2   #  H2   # dT  C jJ.~&T  C jJJKT  C jJ&H2  # ,H2  #  dT  C H2  # 4tT  C dfZB  S D##ZB  S Do%%  0"`#0*$   0"`$% H2  # 8!n"n  C "``! " n  C "`X`!" B  "`tB  "` !N"B  "`( B   "`k3B ! "`D B " "`  B # "`UtB $ "`0B % "`kcB & "`O3B ' "`L3B ( "`NYB ) "`M;h * 3 *"`#")& D T*< 53  c"$n 6 c $X99?"`T*<h 7 3 %7"`,4 %T 8 # $8p   $T 9 # #9  #T : # ": d "P ;  "`Hn < C !<"` ,Z  !T = # =px  ZB > S D,H-ZB ? S D\ <] h @ 3 @"`t"( T A # A4X  Z" BB S ZHI1 Z xZ" C S H]I!  H2 D # x,H2 E # fB F s *D,rsH2 G # x,4H2 H # 4fB I s *D,H2 J # x,dH2 K # dfB L s *D,  T R # R< # ZB T S D ! !h U 3 U"`# H2 V # HH2 W # X H2 X # H2 Y # !t"42 Z  !H!H2 [ # X  !42 \ X  !HTB ] C DX HY ITB ^ C D !H !IB _   H B `B  X B aB  X :!;ZB bB S DlX ZB c S DX %T d # dD%0*_ ZB e S D#dD%eT g # g \  h k 3 k"`Tl0* 42 l  $ @ 42 m @ p $ 42 n  @ @ 42 o \ p$ 42 p  \ 42 q  @ 42 r  $ TB s C D TB t C D B u  N O B vB  & B wB  @ ~  B x   v ~ B yB  B $ xb z <G7HAI@/  xb { <GSHrI`  v $ xb |B <ZGIHQI@$8& &  ' B }B   \ ZB ~ S D,4{B S  ??`      <=B7=C_ ` a u v wxyz{|}N7|&o,$4+(#Ht5D%4  _Ref146378530 _Ref146371092 _Hlt146371102 _Ref146371176 _Ref146371411 _Ref146378610 _Ref146379245 _Hlt146371222 _Hlt146371421 _Hlt146371444 _Ref146378759 _Ref146378168 _Ref146378454 _Ref149220859 _Ref149220900 _Ref149228353 _Hlt149228361 _Hlt149228362 _Ref1463856176 E1*C![]]Rp@@@ @ @@557 F1+C"[^^Sp@? ?XccAy %l(s.BD3FI7LL"MRO[[[aHhijzzD{Ɛvvܓ!%**ŘŘǚ˚˚͡ѡѡ}p  "!#$&%'()*+,-./0123456789:;<=>?h.*urn:schemas-microsoft-com:office:smarttagsCity0http://www.5iamas-microsoft-com:office:smarttagsV@*urn:schemas-microsoft-com:office:smarttagsplacehttp://www.5iantlavalamp.com/i!*urn:schemas-microsoft-com:office:smarttagsState0http://www.5iamas-microsoft-com:office:smarttags_*urn:schemas-microsoft-com:office:smarttagscountry-regionhttp://www.5iantlavalamp.com/Z?*urn:schemas-microsoft-com:office:smarttags PlaceNamehttp://www.5iantlavalamp.com/Z>*urn:schemas-microsoft-com:office:smarttags PlaceTypehttp://www.5iantlavalamp.com/ c@?>@@@@@@@@@@@@@@@.@@@@@@@@.@.@!.@!.@!.@!.@.@!.@!.@!.@?@>.@.; V"["33"4$499::;;<<T<Z<_<k<x<|<<<==>>fAjA5D@DDDIIPP/Q8Q