Task Agnostic Continual Learning Using Online Variational Bayes

By Chen Zeno, Itay Golan, Elad Hoffer, Daniel Soudry. 2018

130 citations recorded by the source corpus.

Catastrophic forgetting is the notorious vulnerability of neural networks to the change of the data distribution while learning. This phenomenon has long been considered a major obstacle for allowing the use of learning agents in realistic continual learning settings. A large body of continual learning research assumes that task boundaries are known during training. However, research for scenarios in which task boundaries are unknown during training has been lacking. In this paper we present, for the first time, a method for preventing catastrophic forgetting (BGD) for scenarios with task boundaries that are unknown during training --- task-agnostic continual learning. Code of our algorithm is available at this https URL.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Reinforcement 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 Reinforcement Learning

Citations in the corpus

Toward Understanding Catastrophic Forgetting in Continual Learning (2019, background): “Improved training procedures have also been developed for this type of approximate Bayesian continual learning through the use of natural gradients [10, 47], ﬁxed-point updates [52], and local approximation [6].”
Improving Task-free Continual Learning by Distributionally Robust Memory Evolution (2022, background): “Task-free continual learning (He et al., 2019; Zeno et al., 2019; Aljundi et al., 2019b; Chrysakis & Moens, 2020) instead focuses on the more general case where the data distribution could change”
Learning to Prompt for Continual Learning (2021, methodology): “To evaluate the proposed L2P, we closely follow the settings proposed in prior works [32,55,66], and conduct comprehensive experiments.”
Continual Learning via Local Module Composition (2021, methodology): “Continual learning methods typically address the problem of forgetting through parameter regularization [46, 70, 99], replay [89, 78, 3, 73, 54, 12, 97, 39, 81, 13] or dynamic architectures (and MoEs) [83, 87, 52, 85, 53, 41].”
ACAE-REMIND for Online Continual Learning with Compressed Feature Replay (2021, methodology): “And we compare with online continual learning methods: AGEM [5], BGD [36], GEM [22], GSS-Greedy [3], HAL [4], ER [31], MIR [1], and GMED [14].”
Layerwise Optimization by Gradient Decomposition for Continual Learning (2021, neutral): “The former approaches, such as EWC [22], SI [55], MAS [1] and ALASSO [35] relied on different estimations of parameters importance, which was usually conducted by exhaustive search [42] or variational Bayesian methods [33, 56].”
New Insights on Reducing Abrupt Representation Change in Online Continual Learning (2021, methodology): “Lastly, Zeno et al. (2018) uses a logit masking related to our method but their context is based on the multi-head setting, and does not consider replay based methods, where learning across tasks occurs.”
Posterior Meta-Replay for Continual Learning (2021, methodology): “In this context, it is important to note that unlike non-Bayesian CL methods, our approach can utilize any type of online prior-focused method (such as BGD [81]) to also learn within tasks in a non-i.i.d. manner.”
Bayesian Structure Adaptation for Continual Learning (2019, methodology): “A natu-ral way to accomplish this is to train a model using online (at task-level) Bayesian inference, whereby the posterior of the parameters learned from task t serve as the prior for task t + 1 (Nguyen et al., 2018; Zeno et al., 2018).”
Continual Unsupervised Representation Learning (2019, background): “Zeno et al. (2018) propose a Bayesian task-agnostic learning update rule for the mean and variance of each parameter, and demonstrate its ability to handle ambiguous task boundaries.”
Three scenarios for continual learning (2019, methodology): “We refer to [12] for a recent insightful treatment of the paradigm without well-deﬁned task-boundaries.”
Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (2020, background): “BGD [85] updates the posterior in closed form and that does not require a task-ID.”
Online Fast Adaptation and Knowledge Accumulation: a New Approach to Continual Learning (2020, methodology): “On the other hand, task-agnostic methods such as Aljundi et al. (2017); Zeno et al. (2018); He et al. (2019) optimize for the ﬁnal performance of the model and thus resort to mechanisms that alleviate catastrophic forgetting.”
La-MAML: Look-ahead Meta Learning for Continual Learning (2020, methodology): “To test our LR modulation against an alternative bayesian modulation scheme proposed in BGD [31], we deﬁne a baseline called Meta-BGD where per-parameter variances are modulated instead of LRs.”
Unsupervised Progressive Learning and the STAM Architecture (2019, methodology): “Other CL variations replace task identiﬁers with task boundaries that are either given [Hsu et al. , 2018] or inferred [Zeno et al. , 2018].”