Human, Animal, and Machine Learning: Experiment and Theory
Abstract: I'll be giving an general overview of my research and research interests. This includes topics in perceptual learning - in particular how task characteristics are primary determinants of the specificity/transfer that can be expected in perceptual learning paradigms - and also how learning that transfers widely enough to be practically useful necessarily requires learning at higher levels of a hierarchical system than at the level of the individual task (using action video games as a model "training" paradigm that teaches at these higher levels, i.e. promotes "learning to learn"). I'll also be discussing work on what we've dubbed "aspiration" - which can be conceptualized as a belief about the "richness" of a given learning environment or "how much there is to learn". These beliefs directly influence the amount of energy one should invest in learning (exploring) and thus how much one actually learns.
Abstract: In probability learning (learning a probabilistic association between one event and another), the conventional wisdom is that learners conform to one of two strategies: a probability matching strategy or a more optimal probability maximizing strategy. Figuring out when people do which one, and why, and how it applies to various domains of learning, has been a critical issue in psychology for over 50 years. In this talk I will review animal and human probability learning and discuss its implications for human language acquisition. A review of this literature in fact suggests that characterizing behavior as the difference between matching and maximizing may not be correct. I will argue that there is in fact no evidence for probability matching, and any deviations from maximizing stem from two factors: flaws in experimental design and data interpretation, or from participants adopting ineffective strategies in the search for a true optimal strategy. This has implications for researchers who are trying to understand the nature of probability learning, and also for how actual probability learning data bears on claims that probability learning has explanatory power over other mechanisms of learning and representation.
Abstract: We study the empirical strategies that humans follow as they teach a target concept with a simple 1D threshold to a robot. Previous studies of computational teaching, particularly the teaching dimension model and the curriculum learning principle, offer contradictory predictions on what optimal strategy the teacher should follow in this teaching task. We show through behavioral studies that humans employ three distinct teaching strategies, one of which is consistent with the curriculum learning principle, and propose a novel theoretical framework as a potential explanation for this strategy. This framework, which assumes a teaching goal of minimizing the learner's expected generalization error at each iteration, extends the standard teaching dimension model and offers a theoretical justification for curriculum learning.
Abstract: Chuck Kalish will be speaking about a project that is the focus of a recent grant submission to IES. The grant proposes to investigate some puzzling failures of transfer in simple learning tasks, including categorization and inference and basic arithmetic, in children and adults. The working hypothesis is that certain instructional practices might promote learning of a discriminative model, in which people learn to map from a feature space to a category label but do not learn much about the distribution of features given the category label. Other practices, in contrast, may promote learning of a generative model, in which the probability distribution over any subset of features can be computed conditioned on the presence of any other subset of features. This idea has been demonstrated in adults inference tasks; we propose to assess (a) whether similar effects are observed in children and (b) whether similar phenomena may be observed in a quite different domain, namely math.
We are beginning to pilot some new experiments, and Chuck will speak about some prior experiments; but there are many outstanding questions especially regarding how best to think about situating arithmetic problems in a multidimensional feature space, and about the relationship between generative and discriminative models more generally. Chuck will be giving a general overview of these issues with hopes that we can have a fruitful discussion that will help us plan the pilot studies.
Abstract: Population games provide a general model of strategic interactions among large numbers of agents; highway congestion, multilateral externalities, and natural selection are among their many applications. To model the dynamics of behavior in population games, we introduce decision protocols, which provide explicit stochastic descriptions of how individual agents make decisions. When the number of agents is large enough, the evolution of aggregate behavior can be described by solutions to ordinary differential equations. We discuss classes of population games in which these evolutionary dynamics lead to equilibrium play, we consider simple examples in which cycling and chaos can arise, and we explain how natural decision protocols can generate potent equilibrium selection results.
Abstract: Psychologists and philosophers who pursue dynamical systems approaches to cognition often see themselves as committed to views that, I shall argue, do not follow from the dynamical perspective. More specifically, dynamicists often take their approach to suggest that minds (or cognition) extend beyond the brain, into the body and even the world. No such conclusion follows from a dynamical perspective. A second position to which dynamicists often believe themselves committed is anti-representationalism. Confusion over the meaning of representation might seem to abet this skepticism, but on a reasonable understanding of the nature of representation, anti-representationalism is implausible.
Abstract: Some of the most influential early work in visual attention concerns the phenomenon of "popout": when deciding whether a horizontal bar is present in a field of vertical distractors, the time needed for a human being to detect the bar is constant regardless of the number of distractors. The standard explanation of this phenomenon is that, when searching for a "primitive" visual feature (ie, not a conjunction of features), people can search all spatial locations in the field in parallel. More recent work in this paradigm has shown that "popout" does not occur when the target and distractors are quite similar (e.g. finding a 15-degree-off-vertical target amongst vertical distractors). Instead, response times grow longer with more distractors.
Both of these patterns contradict the predictions of models of visual saliency. Such models are intended to capture the extent to which parts of the visual field exogenously capture attention. The central idea formalized in such models is that regions of the visual field that carry more information will automatically draw attention. Saliency models applied to neurally-inspired codings of natural scenes do a fairly good job at predicting which elements of the scene will draw a person's attention. When applied to stimuli of the kind used in demonstrations of "popout," however, the models either fail completely or make the seemingly incorrect prediction that people should be *faster* to detect a target when it appears amongst more distractors. /
We tested this prediction in a variant of the standard popout task. Rather than asking people to decide if a bar of a given orientation was present in the display, we instead asked them to judge if all the bars were oriented in the same way or if one was different. Like the standard task, this task requires participants to identify an oddball target amongst similar distractors; but unlike the standard task, the participants do not need to decide if the oddball matches a pre-specified orientation. To our surprise, human behavior conformed to the saliency model predictions: participants were faster and more accurate when the target appeared amongst more distractors. The models failed to match human behavior, however, in another respect. Specifically, the models predict that the facilitating effect of more distractors should be largest when the target and distractors are very dissimilar. We observed the opposite effect: more distractors had an especially large facilitating effect when the target and distractors were quite similar.
I will present the motivation for the work, two different model approaches to saliency including a novel one developed by Nikhil Rao and Rob Nowak, and some of the empirical data we have collected. I am very interested in getting feedback on the work generally, and in particular to brainstorm alternative approaches that might allow us to understand discrepancies between the model and empirical data.
HAMLET mailing list
Tim Rogers (firstname.lastname@example.org), David Devilbiss (email@example.com), Chuck Kalish (firstname.lastname@example.org), and Jerry Zhu (email@example.com) (Add 'u' to the addresses)