search

RHIT-XPrize / rhit-xprize-neural-network

The neural network (and associated training programs) to be used for the Rose-Hulman AI XPrize robot.
2 stars 0 forks source link

Investigate removing "None" nodes. #22

Closed kgroble closed 6 years ago

kgroble commented 6 years ago

We currently have "None" outputs for the neural networks which determine color and letter. Investigate removing these and interpreting low probabilities across the board of concrete results as "none".

rbubulka commented 6 years ago

while it would be pretty simple to just take out the None node and the softmax to adapt to result in low probabilities, the real gain in doing so would be from implementing outlier detection and novelty learning which I do not believe we have time to properly implement with an understanding of what we are doing. Here is two links to the scikit-learn documentation for outlier detection and the commit where they first started implementing it, scikit learn is a tensorflow intermediary similar to keras.

http://scikit-learn.org/stable/modules/outlier_detection.html https://github.com/scikit-learn/scikit-learn/pull/10700/files

rbubulka commented 6 years ago

I believe the best course of action is to ignore this route for now and hand it off to the next team