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

Joel Joseph, Alex GuarXiv 2021 · OpenReview

Abstract

The Continual Learning (CL) problem involves performing well on a sequence of tasks under limited compute. Current algorithms in the domain are either slow, offline or sensitive to hyper-parameters. La-MAML, an optimization-based meta-learning algorithm claims to be better than other replay-based, prior-based and meta-learning based approaches. According to the MER paper [1], metrics to measure performance in the continual learning arena are Retained Accuracy (RA) and Backward Transfer-Interference (BTI). La-MAML claims to perform better in these values when compared to the SOTA in the domain. This is the main claim of the paper, which we shall be verifying in this report.

Introduction

In machine learning, standard neural networks suffer from catastrophic forgetting: when trained on a sequence of tasks, new gradients often overwrite weights optimized for previous tasks, causing performance on earlier data to plummet. While humans excel at "continual learning," artificial systems struggle to balance plasticity (learning new things) with stability (retaining old knowledge).

Existing solutions generally fall into three categories: replay-based (storing old data), prior-based (constraining weight updates), or meta-learning (learning to learn). However, many are computationally expensive or require offline access to data. La-MAML (Look-ahead Meta Learning) addresses these issues by providing an optimization-based meta-learning algorithm that is fast, scalable, and operates in a "single-pass" online setting.

Methods

La-MAML stands for Look-ahead Meta-Augmented Maximum Likelihood. It extends the Model-Agnostic Meta-Learning (MAML) framework with several key innovations:

  • Modulated Learning Rates: Instead of a global learning rate, La-MAML learns a per-parameter learning rate updated during training. This allows the model to effectively "freeze" weights critical for old tasks while remaining flexible for new ones.
  • Asynchronous Updates: Unlike synchronous variants, La-MAML uses the same modulated learning rate for both inner-loop (task-specific) and outer-loop (initialization) updates, which significantly improves stability.
  • Sample Efficient Objective: To reduce computation, La-MAML aligns the current task's gradient with the average gradient of old tasks rather than performing the pairwise alignments used in methods like Meta Experience Replay (MER).

Experimental Setup

The reproducibility study evaluated La-MAML across several standard benchmarks, ranging from simple digit rotations to complex visual classification tasks like TinyImageNet-200. The experiments focused on the "single-pass" setting, where the model sees each data point only once.

Table 1: Dataset statistics and task configurations for the continual learning benchmarks.

Key Results

The authors successfully reproduced the primary claims of the original paper regarding accuracy and efficiency.

  • Performance Metrics: Retained Accuracy (RA) was reproduced to within 6% of the original reported values. As shown in the results table, La-MAML consistently outperformed variants like SYNC and C-MAML across different benchmarks.
  • Efficiency: The study confirmed a dramatic speedup over previous state-of-the-art methods. For instance, on the MNIST Rotations dataset, La-MAML reached its target accuracy in approximately 6 minutes, whereas MER required over 30 minutes.
  • Ablation Confirmation: The modulated learning rate mechanism was identified as a critical component; even without the full meta-learning stack, Experience Replay with modulated rates (La-ER) outperformed standard ER by 15-25%.
Table 2: Comparative performance of La-MAML against its variants C-MAML and SYNC in terms of Retained Accuracy (RA) and Backward Transfer-Interference (BTI).

Discussion and Limitations

While La-MAML is a robust solution, the reproducibility effort highlighted several practical challenges:

  1. Hardware Overhead: Certain experiments, such as "MNIST Many Permutations," have high memory requirements (over 12GB RAM) due to the scale of task loading.
  2. Variance: Results can vary by 5-7% depending on the random seed, suggesting a need for further research into stabilizing meta-learning trajectories.
  3. Asynchrony is Key: A significant finding was that synchronous updates ("Sync") consistently underperformed, validating the specific asynchronous design of La-MAML.

In conclusion, La-MAML is a highly effective and efficient algorithm for online continual learning, offering a scalable alternative to computationally heavy meta-learning approaches.

Summary updated

Abstract

The Continual Learning (CL) problem involves performing well on a sequence of tasks under limited compute. Current algorithms in the domain are either slow, offline or sensitive to hyper-parameters. La-MAML, an optimization-based meta-learning algorithm claims to be better than other replay-based, prior-based and meta-learning based approaches. According to the MER paper [1], metrics to measure performance in the continual learning arena are Retained Accuracy (RA) and Backward Transfer-Interference (BTI). La-MAML claims to perform better in these values when compared to the SOTA in the domain. This is the main claim of the paper, which we shall be verifying in this report.
BibTeX
@misc{joseph2021reproducibility,
  author = {Joel Joseph and Alex Gu},
  title = {{Reproducibility Report: La-MAML: Look-ahead Meta Learning for Continual Learning}},
  year = {2021},
  eprint = {2102.05824},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2102.05824},
}
Figure