This project considers the principles underlying the parallel-distributed processing framework (Center Aim 2), including the principles of learning, processing, and representation. It seeks to refine the core principles that form the basis of the efforts to develop explicit computational models of cognitive processes and to understand the implications of these principles and their relationship to principles arising at other levels. The principles are cast at a level similar to Marr's algorithm level, with the proviso that the underlying biological machinery provides constraints and affordances that shape the processes and representations at the algorithmic level. The considering the algorithmic level, therefore, is strongly guided by constraints arising from the biological level Center Aim 3, with lesser, but still important emphasis on constraints arising from the computational level. The effort is divided into three parts. Part 1 focuses on learning and begins at the algorithmic level, proposing an integrated learning algorithm intended to unify supervised and unsupervised learning. It considers computational effectiveness but places greater weight on biological plausibility, in that it is strongly shaped by our increasing understanding of mechanisms of synaptic plasticity in the brain. Part 2 focuses on representation, beginning at the computational level. It considers whether the representations used in the brain (e.g., the receptive field properties of neurons at various levels of visual processing) can be understood as appropriate solutions to the essential computational problem, namely that of inferring the structure in the world from sensory information. Part 3 focuses on the dynamics of processing and is grounded in our growing appreciation of the details of these dynamics as they are observed in real neurons. It considers (a) the implications of these details for behavior and cognition; and (b) whether these implications can be captured at the more abstract level of the parallel-distributed processing framework via suitable reformulation and extension of the principles. All three parts ultimately target the algorithmic level, in that each has implications that may lead to improved statements of the principles of the framework. The work will interface with all of the other projects in this proposal.
Showing the most recent 10 out of 98 publications