Non-Differentiable Supervised Learning with Evolution Strategies and Hybrid Methods

TL;DR

This paper demonstrates that Evolution Strategies can be scaled to large models and effectively combined with gradient-based methods to train models with non-differentiable parameters like sparsity masks, enabling larger and more efficient models.

Contribution

It introduces a hybrid optimization approach combining ES and gradient descent for models with both differentiable and non-differentiable parameters, scalable to large models.

Findings

ES can be scaled to models with millions of parameters.

Hybrid approach is computationally feasible and competitive.

Enables training sparse models from the first step.

Abstract

In this work we show that Evolution Strategies (ES) are a viable method for learning non-differentiable parameters of large supervised models. ES are black-box optimization algorithms that estimate distributions of model parameters; however they have only been used for relatively small problems so far. We show that it is possible to scale ES to more complex tasks and models with millions of parameters. While using ES for differentiable parameters is computationally impractical (although possible), we show that a hybrid approach is practically feasible in the case where the model has both differentiable and non-differentiable parameters. In this approach we use standard gradient-based methods for learning differentiable weights, while using ES for learning non-differentiable parameters - in our case sparsity masks of the weights. This proposed method is surprisingly competitive, and when parallelized over multiple devices has only negligible training time overhead compared to training with gradient descent. Additionally, this method allows to train sparse models from the first training step, so they can be much larger than when using methods that require training dense models first. We present results and analysis of supervised feed-forward models (such as MNIST and CIFAR-10 classification), as well as recurrent models, such as SparseWaveRNN for text-to-speech.