On Anytime Learning at Macroscale

TL;DR

This paper investigates optimal strategies for training deep neural networks on sequentially arriving data, examining when to update models, how to adapt architectures, and how to balance performance with computational constraints.

Contribution

It formalizes the problem of anytime learning at a macroscale for deep networks and empirically explores strategies for data utilization, architecture adaptation, and training timing.

Findings

Optimal waiting time before training on new data chunks varies.

Adaptive architecture strategies can improve performance over fixed models.

Empirical results on vision and language benchmarks demonstrate practical benefits.

Abstract

In many practical applications of machine learning data arrives sequentially over time in large chunks. Practitioners have then to decide how to allocate their computational budget in order to obtain the best performance at any point in time. Online learning theory for convex optimization suggests that the best strategy is to use data as soon as it arrives. However, this might not be the best strategy when using deep non-linear networks, particularly when these perform multiple passes over each chunk of data rendering the overall distribution non i.i.d.. In this paper, we formalize this learning setting in the simplest scenario in which each data chunk is drawn from the same underlying distribution, and make a first attempt at empirically answering the following questions: How long should the learner wait before training on the newly arrived chunks? What architecture should the learner adopt? Should the learner increase capacity over time as more data is observed? We probe this learning setting using convolutional neural networks trained on classic computer vision benchmarks as well as a large transformer model trained on a large-scale language modeling task. Code is available at \url{www.github.com/facebookresearch/ALMA}.