Blog: Edge of understanding

Edge of understanding
Edge of understanding
Date: 2018-Apr-20 17:31:41 EST

I've been helping with a paper at work, and one of the things it talks about is use of a machine learning technique, applied framewise over time, that will threshold (kinda) if/when data becomes significant enough to become an ROI. I feel I'm on the very edge of an insight that the algorithms in play (a CNN), when viewed stepwise (more or less a 90 degree rotation) might take a very different form and be much more efficiently calculated were they revised towards thresholding rather than independent (but stateful - based on data available so far) invocations. It's a strong instinct, but it's just at the edge of my slippery grasp; if I understood a bit more I feel I could figure out a general transformation of algorithms, or at least develop a means to do so, to realign it along these lines.

A lazy version of this would just be to memoise a lot of the intermediate state.