feeding raw data as image to deep learning machine and expecting to create an image for given class

913 Views Asked by At

I know that deep learning is capable of many cool stuff to do with images. the question that i am facing is: is it possible to create a bitmap image of a large data for multiple class and feed it to a deep learning image processing machine, and when trained expect the machine to generate an image for the given class?

for example predicting a sport match; giving an image of the statics of each game as input, and the class would be the name of two teams. so when I enter "New England Patriots - Seattle Seahawks" the AI generate an image that is the prediction of statics of the game.

will doing so help prediction in anyway?

2

There are 2 best solutions below

2
On

This is an open reasearch topic called Generative Adversarial Networks. An explanation of this can be found here: https://www.quora.com/What-are-Generative-Adversarial-Networks.

Example paper usage: https://arxiv.org/pdf/1612.07828v1.pdf?utm_campaign=Machine%2BLearning%2BWeekly&utm_medium=web&utm_source=Machine_Learning_Weekly_9.

Example youtube tutorial: https://www.youtube.com/watch?v=deyOX6Mt_As.

5
On

Is it possible to create a bitmap image [...] for multiple class and feed it to a deep learning image processing machine, and when trained expect the machine to generate an image for the given class?

As Thomas already wrote, what you might want to have a look at is called "Generative Adversarial Networks" (GAN).

However, it is not as simple to train those as to train feed-forward multilayer perceptrons. You can't simply have the class + noise as input and expect it to generate realistic output images. The problem is:

  1. there are many realistic images (32px x 32px, RGB -> 255^(32*32*3) possible images)
  2. you only know for several million images that they are realistic

The solution is to train two networks which play a game: The generator network G, which generates images. And the discriminator network D, which has to decide if the input comes from the generator or from the natural image source.

for example predicting a sport match; giving an image of the statics of each game as input, and the class would be the name of two teams. so when I enter "New England Patriots - Seattle Seahawks" the AI generate an image that is the prediction of statics of the game.

I'm not sure if I understand what you want. However, if you simply what to display something like

"1:0"

as an image ... why should you use machine learning at all for the displaying-part? Why not simply predict the outcome (machine learning) and generate an image the usual way with probably 3 simple lines of code?

by so many good image processing application in deep learning, will it help me to achieve result more quickly than just input data directly

No. This just adds complexity which has to be solved by the image processing algorithms. After "solving" those problems you artificially added to the dataset, it can tackle the problem you could have given it directly.

i was thinking that i can show more important data like goal scores bigger in the image, so it has more impact in future results

If you know what is important and in which relationship, then don't use machine learning.