Target Layer Regularization for Continual Learning Using Cramer-Wold Generator

Mazur, Marcin; Pustelnik, Łukasz; Knop, Szymon; Pagacz, Patryk; Spurek, Przemysław

doi:10.48550/arxiv.2111.07928

Cited by 1 publication

(2 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…2). We follow the domain incremental setup (van de Ven et al, 2020;Kessler et al, 2021;Raghavan & Balaprakash, 2021;Hsu et al, 2018;Mazur et al, 2021;He & Zhu, 2022), where the number of classes is constant and task information is not available, but we also compare against task-incremental methods (Zenke et al, 2017;Shin et al, 2017;Nguyen et al, 2017;Kirkpatrick et al, 2017;Li & Hoiem, 2017;Rao et al, 2019;Sokar et al, 2021;Yoon et al, 2018;Han & Guo, 2021;Chaudhry et al, 2021 in Appendix C Tables 3 and 4, where the task information is available, to give a complete overview of current methods. The different Continual Learning setups are given in Fig.…”

Section: Continual Supervised Learning Scenariosmentioning

confidence: 99%

“…We achieve comparable results to current state-of-the-art approaches (see Tables 1 and 2) on all three supervised learning benchmarks. (Raghavan & Balaprakash, 2021) 98.71 (± 0.06) 97.51 (± 0.05) Target layer regularization (Mazur et al, 2021) 80.64 (± 1.25) Natural continual learning (Kao et al, 2021) 38.79 (± 0.24) Target layer regularization (Mazur et al, 2021) 74.89 (± 0.61 Memory aware synapses (He & Zhu, 2022) 73.50 (± 1.54)…”

Section: Continual Supervised Learning Scenariosmentioning

confidence: 99%

See 1 more Smart Citation

Hierarchically structured task-agnostic continual learning

Hihn

Braun

2022

Mach Learn

View full text Add to dashboard Cite

One notable weakness of current machine learning algorithms is the poor ability of models to solve new problems without forgetting previously acquired knowledge. The Continual Learning paradigm has emerged as a protocol to systematically investigate settings where the model sequentially observes samples generated by a series of tasks. In this work, we take a task-agnostic view of continual learning and develop a hierarchical information-theoretic optimality principle that facilitates a trade-off between learning and forgetting. We derive this principle from a Bayesian perspective and show its connections to previous approaches to continual learning. Based on this principle, we propose a neural network layer, called the Mixture-of-Variational-Experts layer, that alleviates forgetting by creating a set of information processing paths through the network which is governed by a gating policy. Equipped with a diverse and specialized set of parameters, each path can be regarded as a distinct sub-network that learns to solve tasks. To improve expert allocation, we introduce diversity objectives, which we evaluate in additional ablation studies. Importantly, our approach can operate in a task-agnostic way, i.e., it does not require task-specific knowledge, as is the case with many existing continual learning algorithms. Due to the general formulation based on generic utility functions, we can apply this optimality principle to a large variety of learning problems, including supervised learning, reinforcement learning, and generative modeling. We demonstrate the competitive performance of our method on continual reinforcement learning and variants of the MNIST, CIFAR-10, and CIFAR-100 datasets.

show abstract