In explanation-based learning (EBL), domain knowledge is leveraged in order to learn general rules from few examples. An explanation is constructed for initial exemplars and is then generalized into a candidate rule that uses only the relevant features specified in the explanation; if the rule proves accurate for a few additional exemplars, it is adopted. EBL is thus highly efficient because it combines both analytic and empirical evidence. EBL has been proposed as one of the mechanisms that help infants acquire and revise their physical rules. To evaluate this proposal, 11- and 12-month-olds ( n = 260) were taught to replace their current support rule (that an object is stable when half or more of its bottom surface is supported) with a more sophisticated rule (that an object is stable when half or more of the entire object is supported). Infants saw teaching events in which asymmetrical objects were placed on a base, followed by static test displays involving a novel asymmetrical object and a novel base. When the teaching events were designed to facilitate EBL, infants learned the new rule with as few as two (12-month-olds) or three (11-month-olds) exemplars. When the teaching events were designed to impede EBL, however, infants failed to learn the rule. Together, these results demonstrate that even infants, with their limited knowledge about the world, benefit from the knowledge-based approach of EBL.

Although numerous loop optimization techniques have been designed and deployed in commercial compilers in the past, virtually no common experimental infrastructure nor repository exists to help the compiler community evaluate the effectiveness of these techniques. This paper describes a repository, LORE, that maintains a large number of C language for loop nests extracted from popular benchmarks, libraries, and real applications. It also describes the infrastructure that builds and maintains the repository. Each loop nest in the repository has been compiled, transformed, executed, and measured independently. These loops cover a variety of properties that can be used by the compiler community to evaluate loop optimizations using a broad and representative collection of loops. To illustrate the usefulness of the repository, we also present two example applications. One is assessing the capabilities of the auto-vectorization features of three widely used compilers. The other is measuring the performance difference of a compiler across different versions. These applications prove that the repository is valuable for identifying the strengths and weaknesses of a compiler and for quantitatively measuring the evolution of a compiler.

We present a novel ensemble architecture for learning problem-solving techniques from a very small number of expert solutions and demonstrate its effectiveness in a complex real-world domain. The key feature of our “Generalized Integrated Learning Architecture” (GILA) is a set of heterogeneous independent learning and reasoning (ILR) components, coordinated by a central meta-reasoning executive (MRE). The ILRs are weakly coupled in the sense that all coordination during learning and performance happens through the MRE. Each ILR learns independently from a small number of expert demonstrations of a complex task. During performance, each ILR proposes partial solutions to subproblems posed by the MRE, which are then selected from and pieced together by the MRE to produce a complete solution. The heterogeneity of the learner-reasoners allows both learning and problem solving to be more effective because their abilities and biases are complementary and synergistic. We describe the application of this novel learning and problem solving architecture to the domain of airspace management, where multiple requests for the use of airspaces need to be deconflicted, reconciled, and managed automatically. Formal evaluations show that our system performs as well as or better than humans after learning from the same training data. Furthermore, GILA outperforms any individual ILR run in isolation, thus demonstrating the power of the ensemble architecture for learning and problem solving.

In order for a machine learning effort to succeed, an appropriate model must be chosen. This is a difficult task in which one must balance flexibility, so that the model can capture the complexities of the domain, and simplicity, so that the model does not overfit to irrelevant characteristics of the training data. The optimal model is not only a function of the task to which it is applied, but also the amount of training data available. Copious training data can justify a complex model that includes many of the “true” domain interaction. But when training data is limited, additional simplifications are necessary. Traditional model selection techniques, that require fitting each of a number of hypothesized models to the training data before selecting one, apply in theory, but are not feasible when the number of possible models is large. In this thesis, we describe steps in a new direction for automatically adapting model flexibility. Our approach leverages prior knowledge of two forms: 1) Qualitative knowledge statements, which describe positive and negative relationships between domain variables, and 2) Structural metadata, which provide categorical assignments for each training instance. In our approach, this prior knowledge is used to implicitly construct a large space of alternative well-formed models. A model adaptation procedure then utilizes the training data to conduct a directed search through the space of possible models. The search requires that relatively few models be fit to the data. Thus, the search is efficient and the risk of overfitting in the model selection process is minimized. We demonstrate our approaches on a variety of machine learning tasks, including military airspace safety prediction, planning operator construction, sports prediction, and document sentiment analysis.

We present a novel topic modeling approach to sentiment analysis for documents organized into hierarchical categories. In our approach, positive, negative, and subject matter topics are learned and used to infer document labels. A Markov chain Monte Carlo model procedure adapts the number and structure of topics based on a minimum description length objective function. We apply our approach to Yelp.com business reviews and Amazon.com book reviews and demonstrate that 1) the model adaptation procedure selects a high quality model from the space of alternatives, and 2) the resulting model performs well relative to state of the art regression and topic modeling approaches.

• EM algorithm is a general iterative method of maximum likelihood estimation for incomplete data • Used to tackle a wide variety of problems, some of which would not usually be viewed as an incomplete data problem

Explanation-based Mechanisms for Learning: An Interdisciplinary Approach Michelene Chi (mtchi@asu.edu) Gerald DeJong (mrebl@illinois.edu) Department of Psychology, Arizona State University Department of Computer Science, University of Illinois at Urbana-Champaign Cristine Legare (legare@psy.utexas.edu) Tania Lombrozo* (lombrozo@berkeley.edu) Joseph Jay Williams* (joseph_williams@berkeley.edu) Department of Psychology, University of Texas At Austin Department of Psychology, University of California at Berkeley Abstract The significant role of explanation in learning and generalization is ubiquitous and well documented: explanation promotes student learning in educational settings, drives conceptual development in young children, is accorded a central role in theories of conceptual representation, and has a long history in artificial intelligence. Despite this, relatively little is known about the precise mechanisms that underlie explanation’s effects, and there is a paucity of discourse between the disciplines of cognitive science that study explanation. This interdisciplinary symposium brings together key researchers from education, development, cognitive psychology and computer science to synthesize the progress from these disciplines, forging connections between ongoing research programs to identify promising future directions. Keywords: explanation, self-explanation, understanding, learning, transfer, generalization, explanation-based learning Investigating the processes that foster learning and generalization is one of the core questions in cognitive science: how does a child, adult, or computer acquire knowledge and understanding that is readily generalized to novel situations, rather than lying inert? Explanation has been repeatedly shown to play a key role across a range of core cognitive science disciplines, including education, cognitive development, cognitive psychology, and artificial intelligence. In real-world educational contexts, researchers have documented a self-explanation effect: whether studying math, physics or biology, learners who generate explanations for what they are learning (spontaneously or in response to prompts) are more likely to acquire accurate concepts and transfer their knowledge to novel problems and situations (Chi et al, 1994). Developmentally, generating explanations for the behavior or reasoning of others has been shown to foster conceptual change in young children’s understanding of theory of mind and number conservation (Wellman & Liu, 2006; Siegler, 2002). Research in cognitive psychology also invokes explanation as a powerful force in conceptual representation and revision, even when explanations are not explicitly solicited. In particular, theories of conceptual representation accord a central role to explanation in understanding how concepts and causal relationships between concepts are represented, especially in knowledge-rich domains (e.g. Murphy & Medin, 1985; Lombrozo, 2009). Understanding how explaining promotes learning can thus shed light on concept acquisition and causal learning, and in particular the role of prior knowledge in these processes. Finally, work on explanation-based learning in artificial intelligence and machine learning has isolated different aspects of explanation and formally characterized how these can support generalization (DeJong & Mooney, 1986; Mitchell et al, 1986). However, the developments in these formalisms have not been fully exploited to interpret current educational and psychological work on explanation and learning, nor been directed at understanding current findings and issues in these fields (for a rare exception, see Ahn, Brewer & Mooney, 1992). Despite extensive documentation of the powerful effects of explanation in education and development and the relevance of explanation and learning to current research, little is known about why and how explaining exerts its effects (Lombrozo, 2006). This symposium provides a timely forum for addressing this gap in current theories by bringing together researchers to synthesize empirical findings in education and development, theories and models from artificial intelligence, and investigation using the experimental methodology of cognitive psychology. Michelene Chi synthesizes research in education to give insight into the nature of self-explanation by contrasting it with other learning activities. Cristine Legare provides a developmental perspective on how explanations selectively guide children’s exploration and learning, particularly of causal mechanisms. Joseph Williams and Tania Lombrozo provide a novel account of explanation’s role in generalization in terms of pattern discovery, drawing on philosophical theories and methodology in cognitive psychology. Gerald DeJong presents current machine learning research on how explanations can integrate statistical and logical inference. The constructive nature of self-explanation Michelene Chi conducted the pioneering work on explanation in education with seminal papers on the self- explanation effect. Her talk reviews and synthesizes research in education to shed light on why self-explanation is beneficial, through a comparison to other active, constructive, and interactive learning activities (Chi, 2009). The constructive nature of explanation is proposed as an * Denotes organizers.

Traditional model selection techniques involve training all candidate models in order to select the one that best balances training performance and expected generalization to new cases. When the number of candidate models is very large, though, training all of them is prohibitive. We present a method to automatically explore a large space of models of varying complexities, organized based on the structure of the example space. In our approach, one model is trained by minimizing a minimum description length objective function, and then derivatives of the objective with respect to model parameters over distinct classes of the training data are analyzed in order to suggest what model specifications and generalizations are likely to improve performance. This directs a search through the space of candidates, capable of finding a high performance model despite evaluating a small fraction of the total number of models.We apply our approach in a complex fantasy (American) football prediction domain and demonstrate that it finds high quality model structures, tailored to the amount of training data available.

In this paper we describe the application of a novel learning and problem solving architecture to the domain of airspace management, where multiple requests for the use of airspace need to be reconciled and managed automatically. The key feature of our Generalized Integrated Learning Architecture (GILA) is a set of integrated learning and reasoning (ILR) systems coordinated by a central meta-reasoning executive (MRE). Each ILR learns independently from the same training example and contributes to problem-solving in concert with other ILRs as directed by the MRE. Formal evaluations show that our system performs as well as or better than humans after learning from the same training data. Further, GILA outperforms any individual ILR run in isolation, thus demonstrating the power of the ensemble architecture for learning and problem solving.

Domain knowledge captures an expert's approximate understanding of the world, its objects, and their properties. When available, it should serve to augment the information in a classification learner's training set. But this form of prior knowledge does not easily fit into the statistical learning paradigm. We propose and evaluate the use of phantom examples to remedy this. Our system performs automated model construction and learns generative models for phantom examples that adapt to the need of individual tasks. The approach is validated on the challenging real-world task of distinguishing handwritten Chinese characters. The approach improves learning significantly, provides additional robustness, and works well even though the domain knowledge is imperfect and approximate.

Many successful object detection systems characterize object classes with a statistical profile over a large number of local features. We present an enhancement to this method that learns to assemble local features into Features that capture more global properties such as body shape and color distribution. The system then learns to combine these estimated global features to improve object detection accuracy. In our approach, each candidate object detection from an off-the-shelf gradient-based detection system is transformed into a conditional random field. This CRF is used to extract a most likely object silhouette, which is then processed into features based on color and shape. Finally, we show that on the difficult Pascal VOC 2007 data set, detection rates can be improved by combining these global features with the local features from a state-of-the-art gradient based approach.

In this paper we describe the application of a novel learning and problem solving architecture to the domain of airspace management, where multiple requests for the use of airspace need to be reconciled and managed automatically. The key feature of our “Generalized Integrated Learning Architecture” (GILA) is a set of integrated learning and reasoning (ILR) systems coordinated by a central meta-reasoning executive (MRE). Each ILR learns independently from the same training example and contributes to problem-solving in concert with other ILRs as directed by the MRE. Formal evaluations show that our system performs as well as or better than humans after learning from the same training data. Further, GILA outperforms any individual ILR run in isolation, thus demonstrating the power of the ensemble architecture for learning and problem solving.

This paper focusses on an automated learning/reasoning system for inferring mission safety in airspace operations. We describe a simplified version of a realistic airspace operations scenario that inspired our work. Our domain knowledge about airspace operations and mission safety is expressed qualitatively. We define and describe a way to construct explanations of missions that are translated into a Neural Network representation which is fit to the data and scored. We describe a pruning algorithm to select a greedy-best explanation structure of mission safety. Our experimental evaluation demonstrates the effectiveness of using domain knowledge in learning compared to the standard hidden-layered Artificial Neural Networks. 1 Problem and Significance One of the most important factors in military decisionmaking is safety, i.e., whether a high-valued asset participating in a mission is safe or not, given the enemy threats and our capabilities in that mission. For example, consider the following fictitious but very realistic scenario, simplified substantially here: Intelligence confirms that the militant power in the recently captured Area 6 have escaped to neighboring countries such as Areas 1 and 7 with large numbers of weapons, including SCUD missiles and hand held surface-to-air missiles (SAMs). Despite this intelligence, the exact location of these missiles is still unknown. The Joint Force Air Component Commander (JFACC) directs a large proportion of airpower from Areas 3, 5, and 6 to be devoted to finding and eliminating the threat of SCUD University of Illinois at Urban, Computer Science Department, Siebel Center for Computer Science, 201 N Goodwin Ave, Urbana, IL 61801-2302, USA. Email: {levine,mrebl}@cs.uiuc.edu University of Maryland, Institute for Advanced Computer Studies, College Park, MD 20742, USA. Email: ukuter@cs.umd.edu Analytic Services Inc., 2900 South Quincy St., Arlington, Va 22206, USA. Email: Kevin.VanSloten@anser.org University of Wyoming, Department of Computer Science, 1000 E. University Avenue, Laramie, WY 82071, USA. {anton,derekg,dspears}@cs.uwyo.edu Figure 1: An illustration of the military planning scenario described in the text. The map is based on a screenshot of the map of the well-known board game Risk [Wikipedia, 2009]. missiles. Assets involved in this mission include F-15E Strike Eagles, F/A-18 Hornets, JSTARS, E-2C Hawkeye, E-3A AWACS, and KC-135 tankers. The Rules of Engagement (ROE) require that (1) the coalition forces are not authorized to engage or otherwise cross the border into any neighboring country, and (2) a visual confirmation of any SCUD launched is established before any action – thus, a SCUD hunting mission will only occur during daylight hours or under three-quarter or more moon clear nights. The air defenses of the potential host countries that would interfere with any kind of SCUD hunting mission include multiple shoulder launched surface-to-air missiles, along with the TOR and Hawk (export variant) missile systems. Figure 1 shows some Hawk and TOR missile ranges, and some suspected launcher sights depicted by yellow stars. Hawk missile sights are shown in red, and

When the transition probabilities and rewards of a Markov Decision Process (MDP) are known, an agent can obtain the optimal policy without any interaction with the environment. However, exact transition probabilities are difficult for experts to specify. One option left to an agent is a long and potentially costly exploration of the environment. In this paper, we propose another alternative: given initial (possibly inaccurate) specification of the MDP, the agent determines the sensitivity of the optimal policy to changes in transitions and rewards. It then focuses its exploration on the regions of space to which the optimal policy is most sensitive. We show that the proposed exploration strategy performs well on several control and planning problems.