Neocortical Daisy Architectures and Graphical Models for Context-Dependent Processing
This project is funded by EU IST Future and Emerging Technologies (FP6-2004-IST3) program: ‘Bio-Inspired Intelligent Information Systems’
(Daisy grant number: FP6-2005-015803)
Neocortex has a uniform architecture (the ‘daisy architecture’, DA) consisting of a lattice of patchy populations of pyramidal neurons within cortical areas, and their embedding within inter-areal connections. Our hypothesis is that the DA supports self-organized, context-dependent processing; and that it can be reverse-engineered to provide a major advance in Information Technology by offering novel methods for scalable, distributed, autonomous computation.
We use a combination of neuroanatomical, imaging, theoretical, and hardware engineering methods to evaluate and develop this hypothesis, in a project executed by an inter-disciplinary consortium with record of excellent contributions to research.
We characterize the DA by quantitative neuroanatomical methods and high spatio-temporal resolution imaging of fields of neuronal activity. We explore two candidate models of computation on the DA: graphical models such as Bayesian Networks and Factor Graphs that factor complicated global functions into a product of simpler ones; and Dynamic Link Architectures that encode and recognize objects by dynamic composition of neuronal interactions. We implement examples of these computational styles in hybrid analog/digital CMOS VLSI circuits that can contribute to the ‘End of Moore’s Law Problem’ by demonstrating how existing CMOS technology could be more efficiently deployed than in clocked digital systems.
We consider that computations in DA have a semantic interpretation, in which meaning is expressed by reproducibly linked activity patterns that carry significance in relation to perception and action. We explore this relationship by imaging the activity of cortical neurons as they respond to stimuli chosen for perceptual significance, and by construction of object recognition systems that extract meaningful invariances from examples. These results would be an advance toward incorporating implicit world semantics into computation.