Paper 46, explained

I made an interactive explainer for a conference paper I wrote with a co-author. It is for people who have not read the paper.

A gesture model trained on nine people scored 68.9 percent on a tenth person. Gradient training on five examples per gesture raised that to 79.3 percent, and forward-only retraining reached 81.2 percent within 4,000 forward passes.

It also says what did not work. My formula predicted the cost well for the last layer and only roughly for the whole model, the memory saving was only 1.5 times, and there is no hardware run yet.