One theme studied for a long time within SANS concerns cortical
associative memory function. Previously we have shown that a
recurrent attractor network model with cortical minicolumns as its
computational units is a biologically plausible model of a patch of
cortex. We have demonstrated that such a network is compatible with
cortical dynamics, network organization, and microarchitecture. A
Bayesian-Hebbian learning rule developed at SANS provides a
functional model of synaptic plasticity. The biologically inspired
network architecture and learning rule has further been formalized as
an abstract learning paradigm, BCPNN (Bayesian Confidence Propagation
Neural Network). This can be seen both as an abstract learning model
and a working hypothesis regarding cortical associative memory
function, along the lines suggested by Donald Hebb some fifty years
ago. There is a close collaboration with SICS with regard to
algorithm development and industrial applications of BCPNN.
One theme studied for a long time within SANS concerns cortical
associative memory function. Previously we have shown that a
recurrent attractor network model with cortical minicolumns as its
computational units is a biologically plausible model of a patch of
cortex. We have demonstrated that such a network is compatible with
cortical dynamics, network organization, and microarchitecture. A
Bayesian-Hebbian learning rule developed at SANS provides a
functional model of synaptic plasticity. The biologically inspired
network architecture and learning rule has further been formalized as
an abstract learning paradigm, BCPNN (Bayesian Confidence Propagation
Neural Network). This can be seen both as an abstract learning model
and a working hypothesis regarding cortical associative memory
function, along the lines suggested by Donald Hebb some fifty years
ago. There is a close collaboration with SICS with regard to
algorithm development and industrial applications of BCPNN.