On Anytime Learning at Macroscale
This addresses a practical challenge for practitioners deploying ML in streaming data scenarios, though it is incremental as it builds on existing online learning theory.
The paper tackles the problem of optimally allocating computational budget for training deep neural networks on sequentially arriving data chunks, finding that waiting longer before training on new chunks can improve performance, with gains of up to 5% accuracy on vision tasks and 2% on language modeling compared to immediate training.
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}.