Summary by karpathy 4 years ago
Thanks for the awesome review! I'll break down the reply since commenting seems quite bad in shortscience: We'll definitely improve the paper in v2.

1. We reproduced Siamese as we used the train/test split from MANN. Reimplementing Lake's original algorithm would have been quite involved, but Siamese nets were much closer and simpler to deal with.

2. Note that our convnet uses 28x28 images whereas the original dataset uses 115x115 (IIRC). So there are likely more differences. We used the simpler setup and left large scale experiments for ImageNet, which we deem more adequate for vision.

3. MANN numbers are taken from Table 1. Note that one shot learning is equivalent to 2nd instance (as the first instance is the "training" instance). We used 5th instance for 5 shot learning, but in reality it should be 4 shot learning : )

Your comment:

ShortScience.org allows researchers to publish paper summaries that are voted on and ranked!

Sponsored by: