What this article puts forward AutoAugment The method can automatically select the data enhancement strategy according to the loss value on the verification set .

Problem description

Data enhancement generates more training data based on the existing training sample data , The purpose is to make the expanded training data as close to the real distributed data as possible , So as to improve the generalization ability of the model . Data enhancement is divided into 2 class ： Supervised data enhancement and unsupervised data enhancement . Supervised data enhancement can be divided into single sample data enhancement and multi sample data enhancement ; Unsupervised data enhancement is divided into generating new data and data enhancement strategies . The method of this paper belongs to the 2 Data enhancement strategy in class data enhancement .

AutoAugment A method of automatic data enhancement is proposed , The loss function on the validation set indicates the strategy of data enhancement .

Solve the mind

policy gradient The loss function is introduced into the training method validation set Loss value on
$$\theta =\theta +\alpha {\nabla }_{\theta }(S)\times R$$

Concrete realization

controller Output one strategy（operation, magnitude, probability), Its model structure is a cyclic neural network , One output is the input of the next layer , after 30 individual softmax Layer can be obtained 5 individual sub-policy.

Why it works

Target training to improve the accuracy of downstream tasks , And GAN More interpretable than .

Reference resources

[1] policy gradient Method
[2]Proximal Policy Optimization (PPO) Detailed explanation