CNN Implementation of Bayesian Plasticity for Robust Learning

Bayesian plasticity is presented as an optimal alternative to the delta rule by incorporating uncertainty to guide weight updates. Originally adapted for reservoir computing schemes, Bayesian plasticity showed promise in accelerating the optimization of the mean square error (MSE) function. However, the impact that Bayesian learning rules have in accelerating classification tasks has not been explored. Furthermore, the influence that Bayesian plasticity has on backpropagating weight updates is unclear since reservoir computing only feeds back error to the output layer. Thus, we have simplified and adapted Bayesian plasticity rules to the output layer of non-spiking neural networks and evaluated the impact on training and robustness.

Duke Scholars

Author Hai "Helen" Li Electrical and Computer Engineering