Conference on Cognitive Computational Neuroscience

Predictive coding (PC) is a neuroscientific theory suggesting that the cortex develops a generative model to predict sensory input. Deep neural networks can incorporate PC principles to enhance performance in machine vision tasks or to instantiate cortical models. PredNet, primarily designed for performance improvement, has also been used to interpret the cortex in terms of PC. However, it deviates significantly from the theory, as higher layers predict error unit activity instead of representation unit activity. We propose an alternative deep learning network that adheres more closely to the PC framework by predicting the activity of representation units. Our network generates more realistic neural representations in higher layers while also outperforming PredNet in terms of next-frame prediction. These results call for caution when linking deep learning models optimized for performance to neuroscientific theories and suggest that our model offers a better baseline for combining PC and deep learning.