Methods for maintenance of neural networks in continual learning scenarios
preprintposted on 12.05.2021, 15:24 by Bhasker Sri Harsha SuriBhasker Sri Harsha Suri, Manish Srivastava, Kalidas Yeturu
Neural networks suffer from catastrophic forgetting problem when deployed in a continual learning scenario where new batches of data arrive over time; however they are of different distributions from the previous data used for training the neural network. For assessing the performance of a model in a continual learning scenario, two aspects are important (i) to compute the difference in data distribution between a new and old batch of data and (ii) to understand the retention and learning behavior of deployed neural networks. Current techniques indicate the novelty of a new data batch by comparing its statistical properties with that of the old batch in the input space. However, it is still an open area of research to consider the perspective of a deployed neural network’s ability to generalize on the unseen data samples. In this work, we report a dataset distance measuring technique that indicates the novelty of a new batch of data while considering the deployed neural network’s perspective. We propose the construction of perspective histograms which are a vector representation of the data batches based on the correctness and confidence in the prediction of the deployed model. We have successfully tested the hypothesis empirically on image data coming MNIST Digits, MNIST Fashion, CIFAR10, for its ability to detect data perturbations of type rotation, Gaussian blur, and translation. Upon new data, given a model and its training data, we have proposed and evaluated four new scoring schemes, retention score (R), learning score (L), Oscore and SP-score for studying how much the model can retain its performance on past data, how much it can learn new data, the combined expression for the magnitude of retention and learning and stability-plasticity characteristics respectively. The scoring schemes have been evaluated MNIST Digits and MNIST Fashion data sets on different types of neural network architectures based on the number of parameters, activation functions, and learning loss functions, and an instance of a typical analysis report is presented. Machine learning model maintenance is a reality in production systems in the industry, and we hope our proposed methodology offers a solution to the need of the day in this aspect.