Sparse Coding in a Dual Memory System for Lifelong Learning

Efficient continual learning in humans is enabled by a rich set of neurophysiological mechanisms and interactions between multiple memory systems. The brain efficiently encodes information non-overlapping sparse codes, which facilitates the new associations faster with controlled interference previous associations. To mimic coding DNNs, we enforce activation sparsity along dropout mechanism encourages model to activate similar units for semantically inputs have less overlap patterns dissimilar inputs. This provides us an efficient balancing reusability features, depending on similarity classes across tasks. Furthermore, employ multiple-memory replay mechanism. Our method maintains additional long-term semantic that aggregates consolidates encoded synaptic weights working model. extensive evaluation characteristics analysis show equipped these biologically inspired mechanisms, can further mitigate forgetting. Code available at \url{https://github.com/NeurAI-Lab/SCoMMER}.