Augmentation on the fly - an important tool in the training of neural networks

The most important part of machine learning is data. No matter how good the model and method of training, if the training sample is small or does not describe most of the cases of the real world, it will be almost impossible to achieve high quality work. At the same time, the task of creating training datasets is by no means simple and does not suit everyone, since in addition to the long and exhausting annotation of data by people, additional funding for this process is usually required.

Augmentation, or the generation of new data on the basis of the available data, makes it possible to quite simply and cheaply solve some of the problems with the training set using the available methods. In the case of neural networks, a widespread phenomenon has become to embed augmentation directly into the learning process, modifying the data of each era. However, a very small number of articles focuses on the importance of such an approach and what properties it brings to the learning process. In this article, we will examine what is useful can be extracted from augmentation on the fly, and how critical the choice of transformations and their parameters is within the framework of this approach.

Augmentation: offline or online?

“”, “” “”, . , . , – , , . .

, “” – . : , . , . , “” , . , , AlexNet, ImageNet 2012 . , , : CPU, , GPU. - 1.

 
. 1. -

, , . , , .

: . : , . , , , – , , .

: . , , ( ). , , , ! , , . 2 , .

 
. 2.

, “ ” . , , . , – . , , (, AutoAugment Google, ). – . , , , .

. , : //. – , , . , , (AxialWarp), (RandStructElem) [1] (BrightnessMod). 3.

 
. 3.

, , . , , :

• ;
• , , , (, );
• , . , , 0.9.

( ), , ( ). , , 10% SVHN (Street View House Numbers, , Google Street View), 1. SimpleNet [2] – LeNet , , . , . 4.

1. SimpleNet SVHN

 
. 4. ( ) SVHN

,

, . , , , . , , – .

, ResNet- [3] , ( , ). 5. SVHN. , , , (. 2). 2 – , ! , , .

 
. 5. ResNet- . —

2. ResNet SVHN

. , , – , . , , .

, . “” , . , – , , .

ICMV 2018 (, . ): Gayer, A., Chernyshova, Y., & Sheshkus, A. (2019). Effective real-time augmentation of training dataset for the neural networks learning. Proceedings of SPIE, 11041, 110411I-110411I-8

1. .. / .., .., .., .. // . 2018. . 32. № 3
2. Hasanpour S. H., Rouhani M., Mohsen F., Sabokrou M. Let’s keep it simple, Using simple architectures to outperform deeper and more complex architectures // ArXiv e-prints https://arxiv.org/abs/1608.06037
3. K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. // ArXiv e-prints https://arxiv.org/abs/1512.03385