Thanks! It's a little tricky to derive the nesterov update "backwards", but this looks right to me. It also matches the Caffe implementation, which I will assume was appropriately scrutinized before commit. The slight discrepancy is a minus sign backwards, which stems from the fact that Caffe does the update with a minus sign, keeping the semantics of a gradient throughout, while ConvNetJS simply adds the step to the current parameter vector.

Thanks!