Tracing a naturalistic artificial neural network's adaptation through memory and prediction

An information-theoretic framework has been used to demonstrate that neurons in the visual processing system efficiently distribute resources between storing information contained in past stimuli and predicting future stimuli. The framework can likewise provide a method for evaluating efficiency of allocation of limited computational resources in artificial neural networks (ANNs). We apply the framework to evaluate a feedforward ANN trained to predict visual stimuli. The network mimics the processing of the natural visual cortex by reproducing primary visual cortex receptive fields. We analyzed the efficiency of the network's resource allocation when predicting harmonic oscillation with a Gaussian random forcing impulse, tracing the learning of the ANN through the predictive information plane.