A pattern classifier built from a probabilistic neural network whose radial-basis hidden units are added or removed by a single-pass, hyperparameter-free network-growing algorithm rather than by iterative gradient training. Because units correspond to stored patterns, the network supports continual class-incremental learning and class unlearning by growing or pruning units on demand, reaching far fewer hidden units than a standard probabilistic neural network while matching multilayer-perceptron accuracy and resisting catastrophic forgetting.
This paper proposes a pattern classifier based on a compact-sized probabilistic neural network (PNN) that performs continual, class-incremental learning and also unlearning, the removal of previously…