logo

FIU logo

research
People
publications
Collaborators
software
Main Menu
 
                                                                                                                                     
                                                                                                                                                                                                                                                                                       

                                    object_representation_cleargeneralization_clearcategory_learning_red

                                                            
Category Learning                                                                                        [back]

Human and non-human animals are capable of grouping visual objects into novel categories. A growing body of evidence suggests that at least two category learning mechanisms may be available to people: a procedural learning system, implemented by the circuitry of the basal ganglia, and a rule-based learning system, implemented in the prefrontal cortex.


Learning of representations that support rule learning

rule learningFast, accurate, and generalizable categorization of novel objects depends on the ability to extract some information from the objects that is relevant for a task, while completely ignoring other information irrelevant for the task. Such performance is only possible when the known representation of relevant information (the “relevant dimension”) is independent from the known representation of irrelevant information (the “irrelevant dimension”), so that processing of irrelevant information does not interfere with processing of relevant information. This means that the key to understand fast, accurate and generalizable object category learning is a better understanding of how independent object dimensions are learned from experience and used in categorization tasks. However, very little is known about such processes.

Using multidimensional signal detection theory to measure perceptual independence, we have found that categorization training produces learning of novel separable dimensions. Other experiments suggest that these newly-learned dimensions can be used for rule-based categorization. We have found that subjects using newly-learned face dimensions perform in tests of rule-based category learning as good as subjects trained with the previously-known dimension of gender. Furthermore, high performance is observed after a single one-hour session of training, suggesting that flexible category representations can be learned “on the fly” during category learning.

We are currently exploring what brain areas are associated with learning of independent object representations, using both traditional analyses of fMRI data and novel analyses developed in our previous work on visual object representation using multidimensional signal detection theory.


The role of hubs in the brain networks that support category learning

networkNeurocomputational models explain category learning and automaticity as a function of very specific changes in connectivity between a few brain regions. By themselves, these models do not provide an adequate framework to study the global pattern of network connectivity changes that can be observed during category learning in humans. For this reason, we have recently started to use neurocomputational models in conjunction with network science to study the cognitive neuroscience of category learning. We have studied how changes in the topology of functional networks estimated from fMRI data could explain behavioral changes throughout categorization training. We predicted and confirmed that the basal ganglia and related subcortical structures are focal points of network reconfiguration during procedural category learning, but they lose such role as automaticity develops.

We are currently working on a project that uses network-theoretical measures to determine what cortical regions are important “hubs” of the functional network related to rule-based category learning. We expect that disruption of these areas should be particularly detrimental for rule-based category learning, and we plan to test this prediction in a study using transcranial magnetic stimulation.


  • Soto, F. A., Bassett, D. S., & Ashby, F. G. (submitted). Dissociable changes in functional network topology underlie early category learning and development of automaticity. arXiv 1408.4180

Error-driven learning in object categorization


error learning
What is currently known about the comparative cognition of object category learning suggests that multiple systems for category learning might have originated at different points in evolution. The basal ganglia are homologous structures across vertebrates, suggesting that the procedural learning system thought to be implemented in the basal ganglia, should be similar across species in this group. Our work with pigeons is in line with this hypothesis.

We have developed a theory postulating that object category learning in birds is driven by reward prediction error and implemented in the basal ganglia, using the reinforcement learning mechanisms thought to underlie procedural category learning in humans. Most features of object category learning in pigeons can be explained by this model. Furthermore, the model has generated a number of novel predictions that we have tested and confirmed in object categorization and invariant object recognition studies in both pigeons and people, gathering firm evidence that error-driven reinforcement learning is a “core” mechanism of object categorization, at work across evolutionarily distant species.