Neural networks that perform recognition using generative error may help fill the “Neuro-Symbolic Gap”

Abstract

Abstract Matching the brain’s ability to quickly incorporate new information and have it immediately available for logic and inference remains difficult using feedforward neural network recognition models. Feedforward neural network weights are difficult to modify and are sub-symbolic: they cannot be easily used for logic and reasoning. This work shows that by implementing neural network dynamics differently, during the testing phase instead of the training phase, pattern recognition can be performed using more flexible and symbolically-relevant weights. This advancement is an important step towards the merging of neural-symbolic representations, flexibility, memory, and reasoning with pattern recognition.