A Unifying Bayesian View of Continual Learning

By Sebastian Farquhar, Yarin Gal. 2019

80 citations recorded by the source corpus.

Some machine learning applications require continual learning - where data comes in a sequence of datasets, each is used for training and then permanently discarded. From a Bayesian perspective, continual learning seems straightforward: Given the model posterior one would simply use this as the prior for the next task. However, exact posterior evaluation is intractable with many models, especially with Bayesian neural networks (BNNs). Instead, posterior approximations are often sought. Unfortunately, when posterior approximations are used, prior-focused approaches do not succeed in evaluations designed to capture properties of realistic continual learning use cases. As an alternative to prior-focused methods, we introduce a new approximate Bayesian derivation of the continual learning loss. Our loss does not rely on the posterior from earlier tasks, and instead adapts the model itself by changing the likelihood term. We call these approaches likelihood-focused. We then combine prior- and likelihood-focused methods into one objective, tying the two views together under a single unifying framework of approximate Bayesian continual learning.

Subfields: Probabilistic & Bayesian Methods, 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 Probabilistic & Bayesian Methods

In Meta & Continual Learning

Citations in the corpus

Model Zoo: A Growing Brain That Learns Continually (2021, methodology): “This indicates that replaying data from past tasks is beneﬁcial (Robins, 1995), even if replay may not conform to certain stylistic formulations of continual learning in the literature (Farquhar and Gal, 2019a; Kaushik et al., 2021).”
Posterior Meta-Replay for Continual Learning (2021, methodology): “Prior-focused CL [12] is an alternative Bayesian approach to CL that is commonly used in the literature.”
Optimal Continual Learning has Perfect Memory and is NP-hard (2020, background): “In the same vein, Farquhar & Gal (2018) found that using generative approaches to complement approximately Bayesian procedures improved performance.”
Maintaining Discrimination and Fairness in Class Incremental Learning (2019, methodology): “Other methods construct a generative model, e.g., GANs [8], to generate samples for rehearsal instead of storing old data directly [6, 28, 33].”
Orthogonal Gradient Descent for Continual Learning (2019, background): “, 2018), and understating this phenomenon (Nguyen et al., 2019; Farquhar and Gal, 2019).”
Class-Incremental Learning: Survey and Performance Evaluation on Image Classification (2020, methodology): “VGL [79] introduces Variational Generative Replay, a variational inference”
A Wholistic View of Continual Learning with Deep Neural Networks: Forgotten Lessons and the Bridge to Active and Open World Learning (2020, background): “Both continual generative modeling, Shin et al. (2017); Achille et al. (2018); Farquhar and Gal (2018a); Nguyen et al. (2018); Wu et al. (2018); Zhai et al. (2019), as well as the bulk of class incremental continuous learning works Li and Hoiem (2016); Kirkpatrick et al. (2017); Rebuﬃ et al.…”
Continual Learning using a Bayesian Nonparametric Dictionary of Weight Factors (2021, background): “However, more recent work (Farquhar & Gal, 2018, 2019) suggests that online inference often does not succeed in mitigating catastrophic forgetting in realistic continual learning settings, as methods based solely on online inference rely on the prior capturing everything learned on all previous…”
Overcoming Catastrophic Forgetting by Bayesian Generative Regularization (2019, methodology): “It is also important to compare BGR and VGR, which is a representative algorithm using generative models, in details since both are generative models.”
Continual learning in recurrent neural networks (2020, background): “Some of these have a direct probabilistic interpretation as prior-focused CL methods (Farquhar and Gal, 2018), for which solutions of upcoming tasks must lie in the posterior parameter distribution of the current task (cf. Fig.”
Continual Learning in Recurrent Neural Networks with Hypernetworks (2020, methodology): “Some of these have a direct probabilistic interpretation as priorfocused CL methods [10], for which solutions of upcoming tasks must lie in the posterior parameter”
Unified Probabilistic Deep Continual Learning through Generative Replay and Open Set Recognition (2019, methodology): “Even though VGR [34] uses a single classifier, the researchers trained a separate generative model per task to avoid catastrophic interference in the generator.”
A survey of uncertainty in deep neural networks (2021, background): “[118], [119], [120], theoretic advances in Bayesian learning [121] and beyond.”
Radial Bayesian Neural Networks: Beyond Discrete Support In Large-Scale Bayesian Deep Learning (2019, methodology): “…learning the single-headed environment—where the model has a single output head shared over all tasks and must therefore identify the task as well as the output based on the input—is a much harder task, possibly more reﬂective of continual learning [Chaudhry et al., 2018, Farquhar and Gal, 2018a].”