The core goal of the MIT project is to understand object recognition in naturally cluttered scenes in the cortical ventral visual stream by physiological experiments in primates guided by, and integrated with, a quantitative computational model. The theory and its computer simulations drive the formulation of intedocked hypotheses for different experiments, their planning and the analysis of their results. The plan is: A) to measure invariance and selectivity of IT cells in tasks of object identification with and without natural clutter for different classes of objects to elucidate the neural mechanisms underlying recognition in cluttered, natural scenes; and B) to study the neural correlates of categorization and the relation between categorization and identification suggested by the model.
Our specific aims are: A.1 to determine with single cells recordings the shape tuning of single neurons in IT cortex and their baseline translation invariance for different stimulus classes and clutter conditions; A.2 to measure recognition performance in clutter and correlate it with IT neuron responses; B.1 to verify whether there is a common neural substrate for different recognition tasks, in particular identification and categorization; B.2 to study generalization of categorization in PFC to novel objects and the neural basis of Categorical Perception. The MIT group will collaborate with CalTech in extending the existing quantitative model of recognition -- based on the physiology data from DiCado's lab and the psychophysical data from Caltech --- to include attention and saliency when clutter becomes too severe and the recognition task becomes too difficult for a simple forward flow of information. MIT will also connect the biophysical work of Northwestern and CalTech to the physiology work in IT (DiCarlo) and PFC (Miller) by simulating the implications of properties of the MAX circuitry --- as characterized by intracellular recordings at Northwestern in cat striate cortex --for the recognition model and specifically for its predictions of the selectivity and invariance properties of IT cells in clutter conditions. The corresponding experiments in visual cortex will be done in Jim DiCarlo's and Earl Miller's labs at MIT.
Showing the most recent 10 out of 12 publications
