Gabriel Kreiman
Department of Brain and Cognitive Science, Computation and
Systems Biology Initiative, Center for Biological and Computational Learning
MIT
http://www.mit.edu/~kreiman/Tuesday, March 4, 2006
1065 Kemper Hall
3 :10-4:00 p.m. - Refreshments to follow in 1131
Deciphering Biological Codes for Pattern Recognition
Codes are prevalent throughout biological systems as in the map from the nucleotides in DNA to aminoacids in proteins. Biological codes exhibit emergent properties, non-linear interactions and robustness to perturbations. In this talk, I will discuss how information is represented for pattern recognition in neuronal circuits in the brain focusing on visual object recognition. From all the possible changes in illumination, pose, occlusion, size, color, etc. the visual system can extract a representation of objects that is invariant to these transformations and yet highly selective. I used a biologically plausible, classifier-based read-out technique to investigate the neural coding of selectivity and invariance by studying the responses of a neuronal population in visual cortex. The activity of small neuronal populations over very short time intervals contains surprisingly accurate and robust information about both object identity and category. This information generalizes over object positions and scales, even for novel objects. I will also show that a hierarchical and quantitative model of object recognition could account for the decoding accuracy and robustness of the representation observed in the experimental recordings. This model suggests a quantitative algorithm for the basic initial steps in vision. The methodology, combining machine learning techniques with computational models, provides a general framework for deciphering biological codes.