Simulating Task-Free Continual Learning Streams From Existing Datasets

Abstract

Task-free continual learning is the subfield of machine learning that focuses on learning online from a stream whose distribution changes continuously over time. In contrast, previous works evaluate task-free continual learning using streams with distributions that change not continuously, but only at a few distinct points in time. In order to address the discrepancy between the definition and evaluation of task-free continual learning, we propose a principled algorithm that can permute any labeled dataset into a stream that is continuously nonstationary. We empirically show that the streams generated by our algorithm are less structured than the ones conventionally used in the literature. Moreover, we use our simulated task-free streams to benchmark multiple methods applicable to the task-free setting. We hope that our work will allow other researchers to better evaluate learning performance on continuously nonstationary streams.

Similarity-Based Adaptation for Task-Aware and Task-Free Continual Learning

“…Previous works on task-free CL include (Aljundi et al, 2019a(Aljundi et al, , 2019cRao et al, 2019;Caccia et al, 2020b;Lee et al, 2020;Jin et al, 2021;Pourcel et al, 2022;Wang et al, 2022a;Ye & Bors, 2022;Chrysakis & Moens, 2023;Ye & Bors, 2023). The method by Aljundi et al (2019c) is based on constrained optimization where sample selection is formulated as a constraint reduction problem.…”

“…A memory evolution algorithm is proposed by Wang et al (2022a) in which the memory data distribution dynamically evolves by forcing the memorization process to become gradually harder. Task-free CL streams are simulated by Chrysakis and Moens (2023) from existing datasets via permuting any labeled dataset into a continuously non-stationary stream. The task-free CL algorithm in (Ye & Bors, 2023) aims to improve the compactness levels of the learned structures via a novelty-aware sample selection approach that increases the diversity among the selected memory samples.…”

UER: A Heuristic Bias Addressing Approach for Online Continual Learning

“…3) Adjustability. Other methods [6,10] are difficult to adjust the scale of bias correction, easily limiting the generalization ability of the models. Thus, a meaningful question arises: Is there a more straightforward and efficient approach dot-product logits norm factor angle factor (cosine logits) that can simultaneously meet these three challenges and effectively address the bias issue for OCL?…”