Multi-Label Lifelong Machine Learning Using Deep Generative Replay

Abstract

Lifelong machine learning concerns the development of systems that continuously learn from a set of diverse tasks, incorporating new knowledge without forgetting the knowledge they have previously acquired. One of the main challenges within this paradigm is the stability-plasticity dilemma, which entails balancing a model's adaptability in terms of incorporating new knowledge, corresponding to new tasks, with its stability in terms of retaining previously acquired knowledge (known tasks). Multi-label classification is a supervised learning process in which each instance is assigned multiple non-exclusive labels. When faced with multi-label data, the lifelong learning challenge becomes even more pronounced, as it becomes essential to preserve relations between multiple labels across sequential tasks. This thesis begins with a scoping review that explores the intersection of lifelong learning and multi-label classification.Addressing a major knowledge gap we identified, we introduced Multi-label Deep Generative Replay (MLDGR), a framework for learning from multi-label image datasets across a sequence of tasks with minimal forgetting. As an initial task, the MLDGR approach assigns multiple labels to images. This is achieved by utilizing a convolutional neural network (CNN) for feature extraction from images, which are then input into a base classifier for multi-label classification. In the second stage of our algorithm, selected image samples are transformed into textual representations and stored in a buffer. When learning a new task in the lifelong learning pipeline, the stored textual representations (from prior tasks) are used to generate synthetic images which are then interleaved with real images from the current task. This novel process mitigates forgetting by reinforcing the model's retention of previously acquired knowledge, thereby preventing performance degradation on earlier tasks. Experiments on widely used sequentialized multi-label image datasets demonstrate the effectiveness of our approach and confirm that our algorithm outperforms the state-of-the-art.