My binary vector search is better than your FP32 vectors

Re:My binary vector search is better than your FP32 vectors by heipei on Wednesday March 27, 2024 @05:14PM (#39844688)

To the people dismissing the idea of binarising vectors: Fair criticism, but consider the fact that you can also train a model with a loss function that approaches a binary behaviour, i.e. so that the magnitude per dimension plays an insignificant role and only the sign of the dimension carries information. In that case you can use the binary vector for search and ranking.

[ Parent | on HN ]

Re:My binary vector search is better than your FP32 vectors by mjburgess on Wednesday March 27, 2024 @06:54PM (#39845598)

in practice, learning_rate * grad(L) is mostly about the direction anyway.

[ Parent | on HN ]

Re:My binary vector search is better than your FP32 vectors by niederman on Thursday March 28, 2024 @03:53AM (#39848699)

But the direction is not at all binary, it's a direction in a very high-dimensional weight space.

[ Parent | on HN ]

Re:My binary vector search is better than your FP32 vectors by bjourne on Thursday March 28, 2024 @09:56AM (#39851510)

But the vectors discussed can represent 2^3072 directions. It's close enough.

[ Parent | on HN ]