Niven Achenjang

I've worked on the current neural network code, so I think I have some comments that could be helpful for you working with this codebase (Disclaimer: I haven't read through...

@AtheMathmo if you are still open to allowing someone else to temporarily maintain this project, and are waiting for the right person, I'd suggest @zackmdavis on the relatively simple condition...

I'm not sure that have an enum for different types of layers is the best decision. I feel like that would lead to unnecessarily long match blocks that could be...

I think you're right, `MatrixSlice` would be better than `&[T]`. I have little experience with convolution/pooling layers as well, but I think this approach should be fine. It's based off...

I was thinking about this more, and I realized that this probably wouldn't be a good place to use associated types. If you had some `Activation` layer than computer an...

Only working on `Matrix` would work at first, but we might hit a bump when it comes to conv layers. Convolutions usually evaluate multiple feature maps and output 3D matrices...

That's unfortunate. I tried finding a (messy) [work around](https://play.rust-lang.org/?gist=7f836316fca0f862a1073870b039adbf&version=stable&backtrace=0), but to no avail. I think you're right that this is something we can't avoid. We'll have to not have any...

Ok. I think I'll start trying to implement this, and then submit a PR when I'm far enough along to be critiqued, but before its ready to merge. Does that...

I looked into the issue. I believe to cause is the code running too quickly in release, so the server gets sent too many requests, and can't keep up (The...

I don't know Hyper too well, but that's on the latest version. The bindings for the server use an earlier version where [is seems you have to set this manually](https://docs.rs/hyper/0.9.12/hyper/header/struct.Connection.html#method.keep_alive)....