Continual Learning Using Bayesian Neural Networks

Continual learning models allow them to learn and adapt new changes tasks over time. However, in continual sequential scenarios, which the are trained using different data with various distributions, neural networks (NNs) tend forget previously learned knowledge. This phenomenon is often referred as catastrophic forgetting. The forgetting an inevitable problem for dynamic environments. To address this issue, we propose a method, called Bayesian (CBLNs), enables allocate additional resources without tasks. Using NN, CBLN maintains mixture of Gaussian posterior distributions that associated proposed method tries optimize number needed each task avoids exponential increase involved multiple does not need access past training can choose suitable weights classify points during test time automatically based on uncertainty criterion. We have evaluated MNIST UCR time-series sets. evaluation results show at promising rate compared state-of-the-art models.