We are pleased to present our bold experiment to the habro-community: a calculator based on a neural network. It works as follows: a mathematical expression is converted into an image and fed to the input of a convolutional neural network that generates an image result. The resulting calculator generates images of correct answers without calculating the given expression in explicit form .

The work has already been published on arXiv and will be presented today at the SIGBOVIK conference in audio format. In this post we will share with you the results of our experiment. Motivation and implementation details are also under the cut.

Introduction

- (Generative Adversarial Networks, GANs) , . , .

, adversarial- (GAN). .

. https://yandex.ru/lab/calc_tfjs. , . TensorFlow.js. , , .

. ( ) . , , . , , .

: a — 1920-, b — 1980-, — Windows 3.x, d — , e — .

1920- (. . a) . . , 10 , .

, . , (. . b), . , . , , . . : « . . , , ».

- . , , , .

( ) , 60 , , . , . ( . ): , , .

? . ? , . — , -. — (. . d). , , , (, , ).

, . (. e) — .

. , .

, (, 5 + 2) (, 7). . , MNIST.

UNet- , -. , skip-connections -. UNet. .

, , L1. - MNIST, , MNIST. , GAN perceptual. VGG, MNIST.

. , .

, , . –99 99 . , .

. , , . .

:

, . SIGBOVIK. , , ().

, , . , : , . . , , .

, , . , , . , , . .

, . , . .