There is considerable evidence suggesting that temporal lobe brain areas TE and rhinal cortex of the inferior temporal cortex play important roles in recognizing visual stimuli, and by inference an important role in distinguishing among visual categories such as dogs and cats. Inactivation of rhinal cortex using pharmacological, cooling and ablation interferes with performance in delayed matching and nonmatching to sample (DMS and DNMS) tasks. Neurophysiological recordings identify correlates for recency and familiarity memory in neuronal activity in areas TE, perirhinal cortex and entorhinal cortex. We compared three monkeys in each of 3 groups, those with TE inactivations, those with rhinal (perirhinal + entorhinal) inactivations and three controls on a serial recognition task that did not require simultaneous comparisons between stimuli. In each trial the monkey had to indicate whether the stimulus displayed was new or had been presented earlier within the session. Stimuli were repeated once within the session. The second presentation occurred after no more that 128 intervening stimuli (each of which might be a first or second). The stimulus set had 6000 images of animals in natural scenes. No stimulus was reused within a 30 day period. Control monkeys reliably differentiated between the first and second presentations of a stimulus only had difficulty when the images were separated by 64 or more intervening stimuli. With TE inactivations performance fell from a low level (considerably worse than controls) at short intervals (up to 8 intervening stimuli) to chance levels beyond that whereas rhinal performance was indistinguishable from that of controls. It appears there are some types of recognition memory in which rhinal cortex is not esssential. Visual perceptual categorization is the process through which we assign objects into categories based on some similarity in their appearance. We have shown that monkeys easily learn to discriminate between cats and dogs in a small sample set, and to generalize this ability to never-before-seen exemplars. The monkeys also do this categorization in a set of stimuli with varying degrees of feature ambiguity, where the ambiguity is created by combining or morphing cat-dog pairs with ratios of 100:0, 90:10, 80:20 ..... 0:100). The likelihood of a monkey assigning a cue to a particular category varies sigmoidally with the morph level of the cue. There is a small bias towards one category, presumably because in our testing procedure the rewards are asymmetrical, that is, only one category is rewarded. To investigate how this might happen we developed a reinforcement learning-based spiking neuron model. The model is a neural network with only input and output layers. The input layer neurons all have tuning curves of the same shape (eg. sigmoids, triangles, rectangles, etc.), each neuron with its own parameter values. Each input neuron represents one stimulus group on a continuum of stimuli, i.e., the centers of the turning curves are ordered. The spikes from the neurons of the input layer are summed with adjustable weights by the lone output layer neuron that uses a sigmoidal function to map the summed input into the probability of making the choice leading to reward. The network learns the weights of the inputs onto the single output through reinforcement learning with temporal discounting (in simulations we take temporal discounting over the last 5 trials). The predicted behavioral choices match the experimental findings in producing a sigmoidal performance curve, including the bias toward the rewarded category seen in the behavioral data. When we study the model analytically we can prove a reassuring universality property: the predicted behavioral results are independent of the shape of the neuronal tuning curves. Thus, it appears the form of the tuning curves for neurons for this simple type of reinforcement driven categorization do not play a critical role in learning simple perceptual categorization, providing confidence that the rather simple weighting from inputs to outputs is adequate to explain the results.
Showing the most recent 10 out of 14 publications
