Uni-Logo
You are here: Home Associated Projects Research D7: LeProCoM - Learning and processing of time-varying signals in a laminar-specific cortical microcircuit model
Document Actions

D7: LeProCoM - Learning and processing of time-varying signals in a laminar-specific cortical microcircuit model

D7: LeProCoM - Learning and processing of time-varying signals in a laminar-specific cortical microcircuit model

 

HRI.jpgMarc-Oliver GewaltigD, Jochen Steili, Gerhard SchneiderN, Markus DiesmannQ


D = Honda Research Institute Europe GmbH
iNeuroinformatics Group, University of Bielefeld
N = Computing Center
Q = Riken Brain Science Institute


About the project

In reservoir networks a pool of randomly connected neurons acts as reservoir of non-linear, timevarying functions. A different set of neurons provides a timevarying stimulus, a third set of neurons reads from the reservoir and constructs the desired output function by superimposing the contributions of all reservoir neurons. The connection weights between the reservoir and the read-out neurons are plastic and can be changed to associate an input function with the desired output function. Echo-state networks of analog units and the spiking liquid-state networks are presently the two most important reservoir models. A promissing variation of echo-state networks are the Backpropagation-Decorrelation networks. Three extensions considerably increase the performance: (1) a direct connection between the input and the readout neurons, (2) a learning algorithm which computes the best state of the network for the desired output, and (3) a mechanism which changes the activation function of the neurons to reduce the difference between the achieved output and the desired output.

The project investigates whether layers 2/3 of the cortex can be described as belonging to the class of reservoir networks. In this context a reformulation of the Backpropagation-Decorrelation (BPDC) algorithm for networks of spiking neurons is explored. Biologically plausible processes should be assumed for synaptic as well as intrinsic plasticity. It is an explicit part of the project to develop and implement all models and algorithms in NEST and to evaluate their performance.

 

Funded by Honda Research Institute Europe GmbH

 

Project closed

Personal tools