karpathy / paper-notes Goto Github PK

@karpathy

Hey I've been following your work since a long time now, and I yesterday took this link from twitter feed directing me to this repo containing your notes about the paper. However well written, I wish two more things:

1. Is there a public link to the paper (or some paid journal pointing to it?)
2. Is there any source available?

Sorry for bugging you.

Edit: Forgot to mention, this concerns matching networks only.

You say

I also can't find the numbers reported here in the Siamese Net [11] paper. As far as I can tell in their Table 2 they report one-shot accuracy, 20-way classification to be 92.0, while here it is listed as 88.1%?

They say:

For vision problems, we considered [...] our reimplementation of the Convolutional Siamese Net [11].

the re-implementation and/or training must be different, for example, [11] uses 400 classes, while "Matching Networks" uses 1600 classes (4x rotations).

You say:

I can't find the MANN numbers of 82.8% and 94.9% in their paper [21]; not clear where they come from. E.g. for 5 classes and 5-shot they seem to report 88.4% not 94.9% as seen here. I must be missing something.

MANN paper [21] says:

The network exhibited high classification accuracy on just the second presentation of a sample from a class within an episode (82.8%), reaching up to 94.9% accuracy by the fifth instance

It is also in Table 1