Multi-channel Episodic Memory Building using Recurrent Kernel Machine

Incremental learning is learning of new information without forgetting previous knowledge. Implementation of incremental learning faces the biggest challenge of catastrophic forgetting problem due to stability-plasticity dilemma, algorithms should adapt new information with retaining of previously learned information. So, for providing accurate implementation of incremental learning. We studied the incremental learning process in humans and focused on brain ‘s hippocampus memory involved in learning, information retaining, or recalling. and. By inspiration of human ‘s brain working and architecture we proposed a model in layered architecture, connected hierarchically. First, we develop working memory to automatically extract features ‘s vector of input images using CNN ‘s VGGNet architecture. Second, we develop episodic memory and input feature vector from working memory. Episodic memory is build using recurrent neural network to implement incremental learning with achievement of stability by adjusting weights of neuron and plasticity by adding neuron for unseen input. Also, episodic memory ’s network balance by deleting outlier’s node or edges. having no connection represents no information. Performance of proposed model is evaluated by incrementally learning of KTH dataset’s frame and made comparison with already implementation of IL approaches.