Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

I understand the theory behind neural networks quite well, but am not so clear on how you feed them with images, eg how do you build a network that can process megapixel images of random aspect ratios or audio files of predictable length?

I', trying to get a sense of how much effort would be involved to replicate these results if Google isn't inclined to share its internal tools, to do a neural network version of Fractint as it were, which one could train oneself. I have no clue which of the 30-40 deep learning libraries I found would be best to start with, or whether my basic instinct (to develop a node-based tool in ab image/video compositing package) is completely harebrained.

Essentially I'm more interested in experimenting with tools to do this sort of thing by trying out different connections and coefficients than in writing the underlying code. Any suggestions?



You could try Torch libraries. There are a few examples on how to (almost) replicate some of Google's neural network models on Imagenet.

Check https://github.com/torch/torch7/wiki/Cheatsheet#demos.


Does this work well enough on a modern desktop PC without having access to Google's computing resources?


No idea, but I'm sure their enormous infrastructure helps a lot. Maybe it could be done with some sort of BOINC-type platform.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: