Hacker News Reader

Comment by karterk | Hacker News Reader

What task are you learning the features for? If it's classification, instead of treating it as a two step (feature generation followed by classification) problem, I suggest that you just use a multi-layer neural network with a sigmoid function at the final layer so that you directly predict the output class from raw pixels. This way the feature engineering is taken care of by the algorithm (the weights of the hidden layers).

To give an idea of what I mean, see: http://karpathy.github.io/2015/10/25/selfie/

View on HN

Thank you for the link. I will look into this. Some of the reasons I get the idea I need to incorporate this additional step (feature detection) are:

1) I'm building this primarily for a robot, so my impression is that by building up a set of primitive features, and building up on top of them, will be more efficient for soft real-time use cases.

2) I have done what you suggest (more or less), for example as a test with the MNIST Digit Classification data. I was only getting approx. 95% success rate using this basic technique, which at first glance appears impressive until you realize what is considered "state of the art".

3) My goal is to also incorporate additional layers of data (which I'm under the impression will help improve my overall success rate), such as edge detection, corner detection, and most importantly 3d data (ie. distance from camera based on data extracted from Google Tango), so if I'm correct about #1 where this will be a more efficient calculation, incorporating all the additional data (and extracting features from that as well) will slow things down only incrementally.

1) Once the training is done, the prediction part is simply matrix multiplications - I would think that's probably faster and far easier to parallelize than any other advance feature extraction algorithm you will be performing real time on the image.

2) Did you try using CNN?

I might be grossly wrong here, but I have gone through this path before, but I have come to the conclusion that clumsy hand-crafted feature engineering on images usually performs worse than letting the algorithm figure out. It's much harder to scale and you will end-up coming across a lot of edge cases.

Thank you for the insights. I have not implemented a CNN yet, but is certainly on my radar. In fact other reading the other night pointed me in that direction so I am exploring that at the moment. With regard to scale and/or efficiency of the algorithms, I would guess if anyone is wrong here it would be me :)

BTW, having a look at your link just now. This will be very helpful. Thank you.

Reply on news.ycombinator.com