Continual learning with a predictive coding based classifier

Continual Learning (CL) is the problem of learning multiple tasks sequentially. Several effective CL algorithms using Deep Neural Networks (DNNs) have been developed. However, the problem of reducing the computational requirements of the CL algorithms has not received enough attention. Computationally efficient training methods are important for CL because these models potentially undergo training throughout their lifetime. There is a need to build efficient CL methods that can be used on a broad range of devices. Predictive Coding (PC) is a hypothesis about information processing in the brain. The underlying principle is that the PC model predicts the activity of adjacent layers and updates the model in parallel using local errors between predicted and actual neuron activities, potentially improving the efficiency of CL. This paper proposes a new Continual Learning method using a Predictive Coding based Classifier (CLPC2). CLPC2 trains a PC-based classifier with replay samples generated using a Variational Autoencoder (VAE) or Diffusion (Dif). The performance of CLPC2 is evaluated in three CL scenarios: Class Incremental Learning (Class-IL), Domain Incremental Learning (Domain-IL) and Task Incremental Learning (Task-IL) using the split MNIST, CIFAR-10, and CIFAR-100 datasets. Compared with existing CL methods, CLPC2 achieves higher average classification accuracy in Class-IL and Domain-IL scenarios on MNIST and CIFAR-10 datasets, while obtaining comparable performance on the more challenging CIFAR-100 dataset. The key advantage of the proposed method is the ability to perform training in the classifier using locally computed errors.