Does Preprocessing In Deep Q/Reinforcement Learning Lessen Accuracy?

Question

I've been reading up on deep reinforcement learning such as here:

https://www.nervanasys.com/demystifying-deep-reinforcement-learning/

It will be a while before I understand all the math but that doesn't stop me from using the libraries. Anyway, I know that in a convolutional ANN, if you want to perform learning on images, you have to preprocess the images otherwise the computing power required to compute the neural network is astronomical. Does this degrade the quality of the net in any ways? If so how?

For instance, lets say you had enough computing power to feed the network every pixel of every high quality image in a video stream for the purpose of learning how to reach goals. Would that make the net far more adept at achieving its goal? Would it broaden the types of goals the net could accomplish, possibly giving it the ability to better generalize?

I'm also thinking of this in the context of computer vision where you might have a robot reasoning about its environment in order to learn to perform tasks. It seems that preprocessing the imagery it receives would be akin to handicapping it with extremely poor sight.

Answer 1

By preprocessing the images, do you mean to scale them to a standard size like 256x256 pixels?

By supsampling the image down to a size of 256x256 pixels, you are losing information, so keeping the image in high resolution should give you better results.

But it will also cost you a lot more of computer power, so it's not really worth doing so. In Scaling up Image Recognition the authors claim that by using images of resolution 512x512 instead of 256x256, they achieve a lower error rate. However it is only 0.54% lower, which is not much.

In another tasks I guess the effects of using higher resolution images should be similar, it will let the model see more details hence it will probably be better, but it may not be worth because of the increment in computational power needed.

Note that the authors of the link I provided got caught cheating on ImageNet, so even the decrease in error rate may be real, the score achieved on the ImageNet competition should be ignored.