Visual Images Memory Recall Based on ClassRBM and Free Energy Minimization

Mounting evidence suggests that the brain works in a Bayesian way. Probabilistic inference models based on Bayesian theory have been widely applied in human memory modeling, such as search of associative memory (SAM) and retrieving effectively from memory (REM). However, these lines of memory models can only determine the class of the test images during memory recall. They fail to recall the test visual image, which is referred to as mental imagery in cognitive psychology. In order to address this issue, in this paper, we first propose a memory model based on classification RBM (ClassRBM), in which a three-layer neural network is constructed with 500-800-500 units. The input layer and output layer represent the visual sensory images and corresponding labels, respectively, while the hidden layer stores the encoded visual information. Then we apply free energy minimization principle to train the model, in which we assume that the storage and recall processes of memory in human brain involve free energy minimization. Finally, the learned visual images can be reconstructed by probabilistically sampling the visual units over the hidden units. Comprehensive experiments show that the proposed memory model is capable of recalling images studied and can be applied to the modeling of mental images in the mind.