Lookahead Optimizer: k steps forward, 1 step back

The vast majority of successful deep neural networks are trained using variants of stochastic gradient descent (SGD) algorithms. Recent attempts to improve SGD can be broadly categorized into two approaches: (1) adaptive learning rate schemes, such as AdaGrad and Adam, and (2) accelerated schemes, such as heavy-ball and Nesterov momentum. In this paper, we propose a new optimization algorithm, Lookahead, that is orthogonal to these previous approaches and iteratively updates two sets of weights. Intuitively, the algorithm chooses a search direction by looking ahead at the sequence of fast weights generated by another optimizer. We show that Lookahead improves the learning stability and lowers the variance of its inner optimizer with negligible computation and memory cost. We empirically demonstrate Lookahead can significantly improve the performance of SGD and Adam, even with their default hyperparameter settings on ImageNet, CIFAR-10/100, neural machine translation, and Penn Treebank.

PDF Abstract NeurIPS 2019 PDF NeurIPS 2019 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Stochastic Optimization CIFAR-100 ResNet-18 - 200 Epochs SGD Accuracy 78.24% # 2
Stochastic Optimization CIFAR-100 ResNet-18 - 200 Epochs Polyak Accuracy 77.99% # 3
Stochastic Optimization CIFAR-100 ResNet-18 - 200 Epochs ADAM Accuracy 76.88% # 4
Stochastic Optimization CIFAR-100 ResNet-18 - 200 Epochs Lookahead Accuracy 78.34% # 1
Stochastic Optimization CIFAR-10 ResNet-18 - 200 Epochs Lookahead Accuracy 95.27% # 1
Stochastic Optimization CIFAR-10 ResNet-18 - 200 Epochs SGD Accuracy 95.23% # 3
Stochastic Optimization CIFAR-10 ResNet-18 - 200 Epochs ADAM Accuracy 94.84% # 4
Stochastic Optimization CIFAR-10 ResNet-18 - 200 Epochs Polyak Accuracy 95.26% # 2
Stochastic Optimization ImageNet ResNet-50 - 50 Epochs Lookahead Top 1 Accuracy 75.13% # 1
Top 5 Accuracy 92.22% # 1
Stochastic Optimization ImageNet ResNet-50 - 50 Epochs SGD Top 1 Accuracy 74.43% # 2
Top 5 Accuracy 92.15% # 2
Stochastic Optimization ImageNet ResNet-50 - 60 Epochs Lookahead Top 1 Accuracy 75.49% # 1
Top 5 Accuracy 92.53% # 2
Stochastic Optimization ImageNet ResNet-50 - 60 Epochs SGD Top 1 Accuracy 75.15% # 2
Top 5 Accuracy 92.56% # 1

Methods