Online Continual Learning in Image Classification: An Empirical Survey

By Zheda Mai, Ruiwen Li, Jihwan Jeong, David Quispe, Hyunwoo J. Kim, Scott Sanner. 2021

526 citations recorded by the source corpus.

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, many methods and tricks have been introduced to address this 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 MIR, iCARL, and GDumb and determine which works best at different experimental settings; (2) determine if the best class incremental methods are also competitive in domain incremental setting; (3) evaluate the performance of 7 simple but effective trick such as"review"trick and nearest class mean (NCM) classifier to assess their relative impact. Regarding (1), we observe 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 setting. Overall, this allows us to conclude that MIR is overall a strong and versatile method across a wide variety of settings. For (3), we find that all 7 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.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Radiology, Nuclear Medicine and Imaging, Meta & Continual Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Radiology, Nuclear Medicine and Imaging

In Meta & Continual Learning

Citations in the corpus

A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, background): “In contrast, the latest surveys typically collate only a local aspect of continual learning, with respect to its biological underpinnings [148], [156], [186], [227], specialized settings for visual classification [85], [283], [289], [346], as well as extensions in NLP [37], [206] or RL [214].”
Repeated Augmented Rehearsal: A Simple but Strong Baseline for Online Continual Learning (2022, methodology): “This has been proposed as a trick for online CL to maximally utilize the incoming batch already [Mai et al., 2022, Aljundi et al., 2019]; we investigate whether it is always better than rehearsal with a single iteration.”
DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning (2022, methodology): “However, the task-incremental setting is usually considered over-simplified [3, 33, 36], since task identity is not known at test time in the real world.”
Learning to Prompt for Continual Learning (2021, methodology): “Different from class-incremental, task-incremental learning assumes task identity is known at test time and are often regarded as the simplest setting [34, 38].”
Online Continual Learning on Class Incremental Blurry Task Configuration with Anytime Inference (2021, methodology): “Note that A-GEM performs particularly worse (also observed in (Prabhu et al., 2020; Mai et al., 2021)) as A-GEM was designed for the task-incremental setup and our setting is task-free.”
Online Continual Learning Via Candidates Voting (2021, background): “This setting is more closer to real life learning environment where new classes come in as data streams with limited adaptation time and storage capacity allowed for processing [30].”
Supervised Contrastive Replay: Revisiting the Nearest Class Mean Classifier in Online Class-Incremental Continual Learning (2021, background): “Recent works have revealed that the Softmax classifier and its associated fullyconnected (FC) layer are seriously affected by the class imbalance, which leads to task-recency bias — the tendency of a model to be biased towards classes from the most recent task [39, 37, 24, 57].”
Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (2020, methodology): “We do not consider the online learning setting used in some papers [18], [19] in which each data sample is only seen once.”
Practical Recommendations for Replay-based Continual Learning Methods (2022, methodology): “We select three three of them for our experiments: Split-MNIST [28], Split-CIFAR-10 [33] and Split-TinyImagenet [17].”
Class-Incremental Learning: A Survey (2023, methodology): “We also provide a holistic comparison among different kinds of methods over benchmark datasets, i.e. , CIFAR100 [24] and ImageNet100/1000 [25].”
A TinyML Platform for On-Device Continual Learning With Quantized Latent Replays (2021, background): “task when learning a new one, Continual Learning (CL) has recently emerged as a new technique to tackle the acquisition of new/extended capabilities without losing the original ones – a phenomenon known as catastrophic forgetting [12], [13].”
One Size Fits All for Semantic Shifts: Adaptive Prompt Tuning for Continual Learning (2023, background): “These strategies are detailed in several comprehensive surveys (De Lange et al., 2021; Mai et al., 2022; Mundt et al., 2023).”
Orchestrate Latent Expertise: Advancing Online Continual Learning with Multi-Level Supervision and Reverse Self-Distillation (2024, methodology): “We use two commonly used metrics Average Accuracy ( ACC ) and Average Forgetting ( AF ) to evaluate their performance on benchmarks following [20, 21, 41, 51, 52]: where a i,j is the test accuracy of task i when the model has been trained on task from 1 to j , and T is the total number of tasks.”
CTP: Towards Vision-Language Continual Pretraining via Compatible Momentum Contrast and Topology Preservation (2023, methodology): “Although these methods alleviate forgetting to some extent without storing old samples, they cannot get satisfactory performance in some complex datasets [69] and challenging settings [42].”
Elastic Feature Consolidation for Cold Start Exemplar-free Incremental Learning (2024, background): “…on task datasets during training (Rebuffi et al., 2017; Wu et al., 2019; Masana et al., 2022), while online incremental learning considers continuous data streams in which each sample is accessed just once during training (Lopez-Paz & Ranzato, 2017; Mai et al., 2022; Soutif-Cormerais et al., 2023).”