Reposting my question here, about your reservations w/ the a_i term being dependent on input.

Based on my understanding it doesn't take away from the insight that neural networks are instance-based learners (IMO, a task-relevant compression of the training set), but my follow of the proof is still very hazy and I'd appreciate pushback and insight here.

Thanks for increasing distribution of this paper, I'd love for more people to get eyes on it.