Edward A. Wasserman - US grants

DESCRIPTION (Applicant's abstract): The ability to treat discriminatively different external stimuli as members of a common class is the defining attribute of conceptualization. Prior work has shown that animals as diverse as human beings and pigeons can classify lifelike visual stimuli into natural and artificial categories. The present competitive continuation application aims to see whether the perceptual processes of conceptualization are similar in humans and pigeons. Pigeons will be trained with operant conditioning procedures to discriminate line drawings and computer renderings of natural and artificial stimuli. The pigeons will then be tested with specially modified stimuli that: (1) remove certain portions of the training stimuli, (2) rearrange its component parts, (3) rotate the image in depth. These test stimuli produce highly specific effects in humans, which encourage the view that human object recognition is mediated by a structural description specifying a non-accidental (i.e., viewpoint-invariant) representation of the parts and the relations among those parts--recognition by components. If people and pigeons similarly process visual stimuli, then the results of the planned series of experiments with pigeons should parallel those with people. Empirical convergence would attest to the economy of nature and to the superfluity of language for conceptualization. Empirical divergence would imply that different neurobiological or linguistic mechanisms mediate visual concepts in people and pigeons. In either case, the results of this research project should shed considerable light on the species generality of conceptualization and could yield important clues for how cognitive and computer scientists might more successfully design intelligent systems that are capable of visual pattern recognition. The results could also promote the behavioral analysis of discrimination and generalization processes in learning-disabled or language-deficient humans.

Perceptual similarity models readily explain basic-level categorization. These models propose that, within a category, responses learned to one exemplar transfer to others through primary stimulus generalization. The inadequacy of perceptual similarity as the basis of all truly "conceptual" behavior is noted. Superordinate conceptual categories, for example, often combine basic-level categories encompassing exemplars which are perceptually very different from one another. The proposed studies examine mediated generalization and association as mechanisms of such superordinate categorization. The experiments in the present proposal use a three-step procedure to examine nonsimilarity-based categorization. In a paradigmatic study, diverse exemplars from two basic-level categories (C1 and C2) were associated with one common response (R1) and diverse exemplars from two other basic-level categories (C3 and C4) were associated with another response (R2) in original training. In reassignment training, exemplars from C1 and C3 were associated with new responses (R3 and R4, respectively) and exemplars from C2 and C4 were withheld. In a final test phase, performance of R3 vs. R4 was measured to the exemplars from C1, C2, C3, and C4. If the C1/C2 and C3/C4 combinations were bound into superordinate categories in the first step of training, then subjects should predominately make R3 to the C2 exemplars and R4 to the C4 exemplars. This result was obtained. The proposed studies will examine empirical and theoretical issues concerning mediated generalization and association in superordinate categorization using the three-step procedure. Superordinate conceptual categories are open sets. The word "furniture," for example, is not reserved solely for the familiar lamps and chairs of our own homes, but it is also applied to lamps and chairs we see for the first time. Experiments 1 and 2 of this proposal examine whether superordinate categories created by the three-step procedure are also open sets. Transfer tests with novel exemplars drawn from the component basic-level categories are added to the basic research design. If obtained, transfer of original training to novel exemplars and transfer from novel exemplars to the original training sets will demonstrate that the superordinate categories created in original training of the three-step procedure are indeed, open sets. Another issue to be addressed (in Experiments 3 through 7) concerns the types of training which might bind disparate basic-level categories into one superordinate category. Among the alternatives to be examined are association with: a common delay to reinforcement, reinforcement schedule, contextual stimulus, spatial location, and temporal contiguity. The empirical base established by these studies will give greater perspective on the theories to be considered in of the role of mediation in the formation of superordinate conceptual categories.

Animal Behavior Program

Non-technical Abstract



Title: Entropy and same-different conceptualization

PI: Wasserman, Edward A.

Proposal #: 9904569


Dr. Wasserman and Dr. Young will study the acquisition and application of an abstract
concept by a nonprimate species - the pigeon. Pigeons will be taught to
choose one key when the items of a multi-item display are all the same as
one another and to choose a second key when the items are all different
from one another - a "same-different" discrimination. Previous
experimentation has suggested that the pigeon's discrimination performance
is a function of the variability, or more precisely the entropy, of the
visual displays. In ten experiments, the will seek to better understand the
acquisition of the same-different concept, to explore the generality of the
acquired same-different concept, to study the role of spatial organization
in same-different discrimination learning, and to elucidate the role of
entropy in same-different conceptualization.

This has the potential to disclosing new and important parallels between human and animal cognition, and in demonstrating that even the most advanced forms of cognition can be given precise and rigorous explanations. Documenting that even pigeons are capable of abstraction and that deterministic computational principles may lie at the root of this complex cognitive process should help stimulate neuroscientists to uncover the biological mechanisms of this amazing ability.

Summary The ability to categorize objects and events and to extend this categorization behavior to new instances is foundational to many human activities. We sort the objects and events around us into categories, while still being able to recognize some or all of the individual members of each category. The ability to form categories is especially adaptive, because it allows us to respond appropriately to novel stimuli after experiencing only a few instances from a given category. The same ability should also benefit nonhuman animals. Over the years, animal models have dramatically increased our understanding of the brain mechanisms of human learning and memory, attention and attentional disorders, as well as schizophrenia, aggression, and depression. The proposed research will develop effective pigeon and rat models of human categorization at both behavioral and neural levels. Despite appreciable differences from humans, these animal models will enable direct access to behavioral variables and brain structures that would not otherwise be possible. With both pigeons and rats as experimental subjects, we will systematically explore the role of category structure and task supervision on mastery of the same family of categorization tasks that we will also give to human infants, children, and adults as well as computer models. We will train birds and mammals with tasks that are equated for contingencies of reinforcement, category structure, and task structure, rendering this comparison unique to our collaboration. The bird-mammal comparison will further allow us to explore homologous functions of the hippocampus and striatum as well as analogous functions of the prefrontal cortex and the avian nidopallium caudolaterale. Divergences in categorization behavior between birds and mammals along dimensions that are sensitive to damage to these brain structures will indicate how similar these species are and how they may have diverged through evolution. These results should provide key insights into the behavioral and neural mechanism of categorization with particular relevance to normally and abnormally developing children prior to the emergence of language.