Eric Jang

Eric Jang @ericjang11 · 16h robotics ML practitioner tip: when adding an extra sensor input to your model (e.g. tactile, more history, past image frames), train two baselines along with it (A) random noise (B) zeros instead of your new sensor's values but with the same architecture. If the random/zero baseline makes your model worse, or slows down convergence, it suggests that your sensor fusion architecture / init is suboptimal and may cancel out the benefit or outweigh the effect of the new sensor's information content