Max Planck Institute for Dynamics and Self-Organization -- Department for Nonlinear Dynamics and Network Dynamics Group
Personal tools
Log in

BCCN Sonderseminar

Wednesday, 09.09.2009 17 c.t.
Seminar canceled!

A neural circuit model for categorization and decision making

by Dr. Tatiana Engel
from Department of Neurobiology, Yale University School of Medicine, New Haven, CT, USA

Contact person: Tatjana Tchumatchenko

Location

Seminarraum Haus 2, 4. Stock (Bunsenstr.)

Abstract

Short-term memory is a crucial component to many decision processes, whereby information is maintained over time and interacts with incoming stimulation. Electrophysiological studies in behaving monkeys have identified two neural mechanisms of short-term memory. First, a repeated presentation of a stimulus results in attenuated response. This passive short-term memory in the form of ``repetition suppression'' was predominantly displayed by neurons in area IT of monkeys performing a standard delayed match-to-sample task. The monkeys had to respond to the repetition of a sample in a sequence of test stimuli and performed based on simple detection of any stimulus repetition. The second, active short-term memory was engaged in more complicated tasks, which could not be solved just relying on stimulus repetition, e.g. if distractors could repeat (e.g. in a sequence of stimuli ABBA) or if the match/nonmatch decision was based on abstract category membership of stimuli rather than on their physical similarity. Under these conditions, the sample identity was maintained in persistent firing of some neurons in the monkeys prefrontal cortex (PFC), while two other neural sub-populations showed suppression and enhancement of response to matching stimuli. Based on these observations, we propose a neural circuit model for categorization and match versus nonmatch comparison. The model consists of several interconnected local circuits endowed with three key characteristics. First, spike rate adaptation of single neurons (with a time constant of up to 10s) leads to reduced response to repeated stimuli, i.e. passive repetition suppression. Second, a subpopulation of neurons shows match enhancement as a result of modulatory top-down inputs from a working memory circuit where the identity of a sample stimulus is actively maintained by self-sustained neural firing. Third, category selectivity of neurons is generated through reward-dependent learning. We demonstrate how learning can adjust synaptic weights from the neurons with repetition suppression to the readout network to generate correct match vs. nonmatch decisions using only the passive short-term memory in the standard task. In the ABBA and categorization tasks, the circuit model implements the match vs. nonmatch decision as the competition between two pools of neurons that show match enhancement and suppression. In this case, passive and active short-term memory act in parallel and produce single-cell response patterns that closely match experimental data from PFC. Our model suggests a mechanism of how two forms of short-term memory are used in decision processes in a flexible manner according to task demands.

back to overview