DeepAI AI Chat
Log In Sign Up

In-Loop Meta-Learning with Gradient-Alignment Reward

by   Samuel Müller, et al.

At the heart of the standard deep learning training loop is a greedy gradient step minimizing a given loss. We propose to add a second step to maximize training generalization. To do this, we optimize the loss of the next training step. While computing the gradient for this generally is very expensive and many interesting applications consider non-differentiable parameters (e.g. due to hard samples), we present a cheap-to-compute and memory-saving reward, the gradient-alignment reward (GAR), that can guide the optimization. We use this reward to optimize multiple distributions during model training. First, we present the application of GAR to choosing the data distribution as a mixture of multiple dataset splits in a small scale setting. Second, we show that it can successfully guide learning augmentation strategies competitive with state-of-the-art augmentation strategies on CIFAR-10 and CIFAR-100.


page 1

page 2

page 3

page 4


Improving Generalization in Meta-Learning via Meta-Gradient Augmentation

Meta-learning methods typically follow a two-loop framework, where each ...

Direct Differentiable Augmentation Search

Data augmentation has been an indispensable tool to improve the performa...

Non-Differentiable Supervised Learning with Evolution Strategies and Hybrid Methods

In this work we show that Evolution Strategies (ES) are a viable method ...

Learning to Augment via Implicit Differentiation for Domain Generalization

Machine learning models are intrinsically vulnerable to domain shift bet...

Efficient Optimization of Loops and Limits with Randomized Telescoping Sums

We consider optimization problems in which the objective requires an inn...

Fast Gradient Methods with Alignment for Symmetric Linear Systems without Using Cauchy Step

The performance of gradient methods has been considerably improved by th...

Learning Augmentation Distributions using Transformed Risk Minimization

Adapting to the structure of data distributions (such as symmetry and tr...