Lattice Neural Networks for Incremental Learning

In incremental learning, it is necessary to conquer the dilemma of plasticity and stability. Because neural networks usually employ continuously distributed representation for state space, learning newly added data affects the existing memories. We apply a neural network with algebraic (lattice) structure to incremental learning, that has been proposed to model information processing in the dendrites of neurons. It has been proposed as a mathematical model of information processing in the dendrites of neurons. Because of the operation 'maximum' in lattice algebra weakening the continuously distributed representation, our proposed model succeeds in incremental learning.