4 ms·
But the gigantic synthetic dataset that is used for training is created with plenty of traditional search. So it is all a bit silly but I guess cool none the le
by chvid 2y ago
But the gigantic synthetic dataset that is used for training is created with plenty of traditional search. So it is all a bit silly but I guess cool none the less ...
- chvid 2y agoIf anything it demonstrates the limits of NN. A human brain can learn based on far fewer examples.
- jxy 2y agoNature's evolution algorithm took millions of years to find the architecture and the base model, which then takes decades to be fine tuned to be able to form this opinion.
- amunozo 2y agoIt's a knowledge distillation. You can then use this smaller, more efficient models instead of the larger one.
- chvid 2y agoOr maybe it is just memorizing a very large number of games.
- azakai 2y agoThey address the possibility of memorization in the PDF: > This effect cannot be explained by memorization since < 1.41% of the initial puzzle board states appear in our training set.
- tech_ken 2y agoSeems more like a 'compression' of the large number of games, or even like an approximate 'index' of the database
- alkonaut 2y agoIs this network smaller than stockfish and by what metric is that?
- msoad 2y agoSearched only once. If this can be applied to other knowledge with this efficiency we're onto something