Abstract:Online continual learning for image classification studies the problem of learning to classify images from an online stream of data and tasks, where tasks may include new classes (class incremental) or data nonstationarity (domain incremental). One of the key challenges of continual learning is to avoid catastrophic forgetting (CF), i.e., forgetting old tasks in the presence of more recent tasks. Over the past few years, a large range of methods and tricks have been introduced to address the continual learning problem, but many have not been fairly and systematically compared under a variety of realistic and practical settings. To better understand the relative advantages of various approaches and the settings where they work best, this survey aims to (1) compare state-of-the-art methods such as Maximally Interfered Retrieval (MIR), iCARL, and GDumb (a very strong baseline) and determine which works best at different memory and data settings as well as better understand the key source of CF; (2) determine if the best online class incremental methods are also competitive in the domain incremental setting; and (3) evaluate the performance of 7 simple but effective tricks such as the "review" trick and the nearest class mean (NCM) classifier to assess their relative impact. Regarding (1), we observe that iCaRL remains competitive when the memory buffer is small; GDumb outperforms many recently proposed methods in medium-size datasets and MIR performs the best in larger-scale datasets. For (2), we note that GDumb performs quite poorly while MIR — already competitive for (1) — is also strongly competitive in this very different (but important) continual learning setting. Overall, this allows us to conclude that MIR is overall a strong and versatile online continual learning method across a wide variety of settings. Finally for (3), we find that all tricks are beneficial, and when augmented with the "review" trick and NCM classifier, MIR produces performance levels that bring online continual learning much closer to its ultimate goal of matching offline training. Our codes are available at https://github.com/RaptorMai/online-continual-learning.

Multimodal Continual Learning Using Online Dictionary Updating.

Lifelong Learning For Heterogeneous Multi-Modal Tasks

Deep Vision Multimodal Learning: Methodology, Benchmark, and Trend

Recent Advances of Multimodal Continual Learning: A Comprehensive Survey

Continual Learning Meets Multimodal Foundation Models: Fundamentals and Advances

Visual–tactile fusion material identification using dictionary learning

Continual Learning for Multimodal Data Fusion of a Soft Gripper

Adaptive online continual multi-view learning

A unified multimodal classification framework based on deep metric learning

Multi-Modal Convolutional Dictionary Learning.

Multimodal Task-Driven Dictionary Learning for Image Classification

Dual-Memory Multi-Modal Learning for Continual Spoken Keyword Spotting with Confidence Selection and Diversity Enhancement

Online Continual Learning through Mutual Information Maximization.

Comprehensive Semi-Supervised Multi-Modal Learning.

Learning to Learn Better Unimodal Representations Via Adaptive Multimodal Meta-Learning

Incomplete Multimodal Learning for Remote Sensing Data Fusion

Multimodal Representation Learning by Alternating Unimodal Adaptation

Online continual learning in image classification: An empirical survey

A New Dictionary Construction Based Multimodal Medical Image Fusion Framework.

Vision-Sensor Attention Based Continual Multimodal Egocentric Activity Recognition.

Cross-Modal Alternating Learning with Task-Aware Representations for Continual Learning