Stimulus-specific versus abstract learning: a computational approach

H. G. REY, D. GUTNISKY, B. S. ZANUTTO

Washington, DC USA

Congreso; Neuroscience, 2008. Program No. 790.3/TT18.; 2008

Society for Neuroscience

The capacity to learn stimulus-specific and abstract associations is essential in human cognition. However, the mechanisms involved in the use of these different strategies are unknown. Moreover, the possibility of nonprimate animals having both abilities is still on debate.Conceptualizing contingencies abstractly is an efficient way to save memory resources. In the past, Same/Different (SD) discrimination, a simple form of abstract learning, was thought to have language training as a necessary condition. However, experimental evidence has shown that this task cannot only be solved by humans, but also by chimpanzees, rhesus and capuchin monkeys, and pigeons. In addition, neurophysiological data suggests that neurons in prefrontal cortex of rhesus monkey code abstract rules.We have previously introduced a model for stimulus-specific learning. Here, we propose a minimal model that is able to solve the SD task. The model includes working memory units, SD units and response units. Each different stimulus activates a defined subset of working memory units. The synapses between working memory units and another cluster of units include an adaptation mechanism. This nonlinearity allows the generation of differential signals when the attributes of two stimuli match or mismatch. The resulting smaller activation on same trials is irrespective of the specific features of the stimuli (just the relation among the features of each stimulus). Processing the differential signal for all the features activates different clusters on same and different trials. Response units pool these clusters to provide the correct response (with the particular responses being associated to the SD clusters by reinforcement learning).The model is minimal in the sense that all its components are already described in the literature, although without explicit mention to its relevance to this task. The model also provides a basic building block with which more complicated abstract tasks can be accomplished, such as conceptual matching to sample, learning sequences of stimuli, etc.We also discuss a potential mechanism for multiplexing (or even mixing) the results processed by the stimulus-specific and abstract subsystems. Particularly, we emphasize the potential role of the hippocampal system. Different animals may have different biases towards one of those subsystems but, according to the task demands, they might be able to modify the relative weights for each strategy. This might in turn be important to support the idea that several species (e.g., humans, monkeys and pigeons) present quantitative but not qualitative differences regarding to their ability to learn abstract concepts.