Continual Learning for Classification Problems: A Survey

Abstract

Artificial Neural Networks performs a specific task much better than a human but fail at toddler level skills. Because this requires learning new things and transferring them to other contexts. So, the goal of general AI is to make the models continually learning as in humans. Thus, the concept of continual learning is inspired by lifelong learning in humans. However, continual learning is a challenge in the machine learning community since acquiring knowledge from data distributions that are non-stationary in general leads to catastrophic forgetting also known as catastrophic interference. For those state-of-art deep neural networks which learn from stationary data distributions, this would be a drawback. In this survey, we summarize different continual learning strategies used for classification problems which include: Regularization strategies, memory, structure, and Energy-based models.