La-MAML: Look-ahead Meta Learning for Continual Learning

By Gunshi Gupta, Karmesh Yadav, Liam Paull. 2020

81 citations recorded by the source corpus.

The continual learning problem involves training models with limited capacity to perform well on a set of an unknown number of sequentially arriving tasks. While meta-learning shows great potential for reducing interference between old and new tasks, the current training procedures tend to be either slow or offline, and sensitive to many hyper-parameters. In this work, we propose Look-ahead MAML (La-MAML), a fast optimisation-based meta-learning algorithm for online-continual learning, aided by a small episodic memory. Our proposed modulation of per-parameter learning rates in our meta-learning update allows us to draw connections to prior work on hypergradients and meta-descent. This provides a more flexible and efficient way to mitigate catastrophic forgetting compared to conventional prior-based methods. La-MAML achieves performance superior to other replay-based, prior-based and meta-learning based approaches for continual learning on real-world visual classification benchmarks. Source code can be found here: this https URL

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, 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 Meta & Continual Learning

Citations in the corpus

CoMPS: Continual Meta Policy Search (2021, methodology): “Recent meta-learning based methods have made inroads on the continual meta-learning but they are limited to non-RL problems (K J & N Balasubramanian, 2020; Gupta et al., 2020).”
Flattening Sharpness for Dynamic Gradient Projection Memory Benefits Continual Learning (2021, methodology): “EWC 62.25 ± 1.44 -15.22 ± 1.25 GEM 61.82 ± 0.85 -15.58 ± 1.17 ER 68.31 ± 0.56 -13.91 ± 0.67 La-MAML 75.98 ± 0.60 -10.21 ± 0.90 GPM 74.54 ± 0.36 -7.17 ± 0.51 FS-DGPM 76.96 ± 0.51 -7.45 ± 0.30 Multitask 86.54 ± 1.74 - PMNIST: We first evaluate our algorithm for 20 sequential PMNIST tasks with only 100”
Continual Learning on Noisy Data Streams via Self-Purified Replay (2021, background): “Recently, [77] adopts graphs to represent relational structures between samples, and [25] employs the meta-loss for learning per-parameter learning rates along with model parameters.”
Reproducibility Report: La-MAML: Look-ahead Meta Learning for Continual Learning (2021, neutral): “La-MAML [1] claims to be a better performing, robust and faster algorithm compared to the existing baselines.”
Online Continual Learning in Image Classification: An Empirical Survey (2021, background): “La-MAML [64] proposes an efficient gradient-based meta-learning that incorporates per-parameter learning rates for online CL.”
Reactive Exploration to Cope with Non-Stationarity in Lifelong Reinforcement Learning (2022, methodology): “Other approaches consider meta-learning (Schmidhuber, 1987; Bengio et al., 1992), which uses task-specific inner loops and task agnostic outer loops to learn supportive representations that are helpful across tasks (Javed & White, 2019; Gupta et al., 2020).”
Formalizing the Generalization-Forgetting Trade-off in Continual Learning (2021, methodology): “Despite significant progress, however, these methods [22, 3, 20, 16, 23, 48, 1, 15] are still inherently tilted toward maximizing generalization or minimizing forgetting because they naively minimize the loss function.”
How to Reuse and Compose Knowledge for a Lifetime of Tasks: A Survey on Continual Learning and Functional Composition (2022, methodology): “These same techniques have been used with a meta-learning objective function, whereby the agent trains directly to optimize the network’s feature representation to avoid gradient conflicts between future and previous tasks (Riemer et al., 2019; Gupta et al., 2020a).”
TAG: Task-based Accumulated Gradients for Lifelong learning (2021, background): “Another related work [22] also employs an adaptive learning rate while requiring a small episodic memory.”
Sequoia: A Software Framework to Unify Continual Learning Research (2021, unknown): “, 2017) „ „ X X „ „ X X La-MAML (Gupta et al., 2020) H H X X „ „ X H replay H H X X „ „ X H HAT (Serrà et al.”
Forget-free Continual Learning with Winning Subnetworks (2022, methodology): “Also, † and ∗ denotes results reported from (Deng et al., 2021) and (Gupta et al., 2020) (b), the capacity of Conv4 is greater than that of Conv1, while the Conv4 proportion of reused weights for all tasks is smaller than Conv1.”