|
| home |
| research |
| publications |
| vitae |
| contact info |
Research Overview
A central challenge for all cognitive systems is using past experience to guide behavior in novel situations. Furthermore, the effectiveness of this type of learning can depend critically on how stimuli and concepts are represented, that is, how perceptual information and semantic knowledge are organized. These issues have led me to focus my research at the interface of perceptual representation, conceptual representation, and learning, with emphasis on the following core questions.
I study these questions using an interplay of human behavioral experimentation and mathematical and computational modeling, guided at times by neuropsychological theories. My primary domains of inquiry are category learning, similarity, and sequential decision making, although I have also done research in game theory, working memory, social choice, and pure math (topology).
Sequential effects in category learning
Categories play a fundamental role in cognition, in that they allow people to group past experiences and rapidly generate expectations and responses for new stimuli. A great deal of research has tried to answer the questions of how categories are learned and represented, and how category learning interacts with perceptual representation. In my own work, I have found that sequential effects during learning provide a powerful tool for addressing these issues. First, by separating the effects of recent stimuli and recent feedback, Jones, Love, and Maddox (2006) show how to distinguish perceptual and decisional components of sequential effects. These have been confounded in previous research in categorization as well as other domains such as psychophysical scaling. Second, decisional sequential effects can be used to assess detailed aspects of learning and representation. Jones (in preparation) shows how sequential effects can provide a direct indicator of how categories are represented (i.e., exemplars vs. rules vs. prototypes). Jones, Maddox, and Love (2005) use a similar approach to investigate how selective attention and perceptual representations shift in order to support learning. Third, we have used analyses of sequential effects to demonstrate limitations on Shepard’s influential Universal Law of Generalization, which states that people generalize knowledge from one stimulus to another according to their perceived similarity (Jones, Maddox, & Love, submitted). Specifically, when judgments depend on more than one psychological dimension, generalization depends on the full multidimensional relationship between stimuli in a way that cannot be mediated by any unidimensional similarity function.
Temporal structure in repeated tasks
Information from past experience is not all equal in its relevance to the current situation. Statistical analyses of environmental information sources have shown that more recent information tends to be more relevant and reliable. This has been proposed as a functional basis for decay in memory and other recency effects in judgment. Much of my research in graduate school tested this hypothesis by investigating how people take advantage of temporal structure in repeated tasks, specifically by learning to rely appropriately on recent information. This research included connectionist (recurrent-network) modeling of working memory (Jones & Polk, 2002), experimental work in category learning (Jones & Sieck, 2003), and formal analysis of repeated games such as the Iterated Prisoner’s Dilemma (Jones & Zhang, 2004). Together, these studies showed that people can selectively rely on recent information in a way that adapts to environmental statistics and proposed specific learning mechanisms by which this might take place.
Dissociating short- and long-term contributions to learning
My work to date on sequential effects and learning of temporal structure raises the question of whether processing of recent information relies on the same cognitive mechanisms as long-term learning. That is, there is the intriguing possibility that learning is subserved by separate short- and long-term systems that operate on different types of representations. I am evaluating this new dual-system theory of learning in a series of categorization experiments with Todd Maddox and Brad Love, designed to test for dissociations between how people use recent and more temporally distant information. The general picture emerging thus far is that the long-term system learns over time to associate stimuli to appropriate responses, whereas the short-term system uses a more dynamic strategy based on comparisons between the current stimulus and other recently encountered stimuli. This dynamic calibration strategy can produce surprisingly good performance in many repeating or interactive tasks even in the absence of true long-term learning.The contribution of the short-term system to behavior has potentially wide-ranging implications. First, previous research in category learning has conflated short- and long-term processes, suggesting that many previous findings assumed to be diagnostic of long-term category representations may instead reflect operation of the short-term system. Second, the short-term system may play an especially important role in situations where reliable long-term knowledge is not available. For example, findings of spared category learning in amnesics may be explainable by appeal to this system. Likewise, short-term processing may play a big role in psychophysical scaling and identification with unidimensional stimuli, where it has recently been argued that people lack reliable representations of absolute stimulus magnitudes in long-term memory. In future work, I plan to explore the connections of this theory to models of short- and long-term memory and to other dual-system learning theories (e.g., the COVIS model of Ashby and colleagues) and to investigate its neurological underpinnings.
Relational information in conceptual representation
The efficacy of short-term processing in traditional learning paradigms shows the power of relative information. Rather than classifying stimuli based on their absolute feature values, people can (and often do) solve categorization tasks by relying on comparisons to other recent stimuli. In a related line of research, I have been investigating how relational information contributes to more conceptual-level knowledge representation. In contrast to classical theories that assume concepts are represented in terms of their intrinsic properties (such as wings for bird), several researchers have recently proposed that the meaning of concepts is largely determined by their relationship to other concepts. Jones and Love (in press) tested this idea by investigating the contribution of relational information to similarity. We found that items become more similar when they participate in the same relation (e.g., The knife cuts the bread) and when they play the same role within a relation (e.g., The seal chases the fish; The dog chases the cat). These findings contrast with essentially all previous theories of similarity, which focus exclusively on features or internal relational structure, and indicate that a significant component of concept representations involves their relations to other concepts.The influence of relations on similarity also has implications for models that induce conceptual structure from verbal input, such as Latent Semantic Analysis. Although these models can extract a great deal simply from word co-occurrence statistics, they fail to pick up on role-based properties due to their insensitivity to grammatical structure. In response to this shortcoming, we developed a new model that operates on syntactic trees instead of unordered input and that correctly predicts the relational similarity effects found in our empirical investigations. In future work with Ken McRae (University of Western Ontario) we plan to apply the model to large corpora to investigate what emergent capabilities its sensitivity to relational information may provide on richer input sets.
Representation of integral dimensions
Stimuli defined by integral dimensions, such as colors, tend to be processed in a holistic fashion rather than by analyzing their component dimensions. This suggests that the perceptual representations of spaces defined by integral dimensions are relatively unstructured. In a recently completed project, Jones and Goldstone (in preparation) investigated how much structure the representations of integral dimensions have and how those representations change when, through training, people learn to differentiate dimensions that were initially integral. Dimension differentiation is thought to be a central process in children’s perceptual development and development of selective attention. We proposed two models that differ in the amount of geometrical structure present in integral-dimension perceptual spaces and tested their predictions on a series of training and transfer categorization tasks. Our results showed that representations of integral dimensions contain a surprisingly rich structure, one that plays an important role in dimension differentiation.
Future work: Learning and representation in temporally extended tasks
In current work, I am extending models of category learning to more realistic, dynamic tasks. Unlike categorization and most other laboratory paradigms, natural tasks generally require the decision maker to balance immediate and delayed consequences of actions and to coordinate sequences of actions to achieve desired goals. I believe there is great potential for synthesizing research in psychology and machine learning to address this problem. Although cognitive modeling has not made much progress on sequential decision making, there are powerful algorithms based in control theory, collectively known as reinforcement learning (RL) algorithms, that have greatly improved the ability of autonomous systems to learn complex, dynamic tasks. Recent neurological work with humans and other animals has found evidence for the basic learning mechanism behind these algorithms, making them strong candidates for integration into cognitive-level models. Conversely, a major impediment to RL algorithms in machine learning is that they depend heavily on how the modeler hand-codes how stimuli are represented, whereas research in human learning, category learning in particular, has uncovered powerful mechanisms by which people adapt their internal representations to suit the task and to support rapid learning.The model I am developing combines flexible representational learning mechanisms developed for modeling human category learning (Love & Jones, 2006) with learning of action sequences via RL. By combining the strengths of these two fields, it has the potential to make significant progress in both. There has also been a recent surge of interest in RL within systems neuroscience, and I intend to pursue connections to that work as well. Recently I was awarded a small grant from Applied Research Associates (a private company I have been consulting for) to apply the model to benchmark problems in the machine learning literature. Initial simulations show the model is able to learn sequences of actions to achieve long-term goals and to adjust its attention and the complexity of its internal representation to fit the task. With Brad Love, I have also written a grant proposal for a series of experiments on humans testing some of the model’s key predictions. These predictions for how people learn complex tasks have practical implications for education and training of military and other expert personnel.