Atlas LSH neural networks With locality sensitive hashing…
1 reactions
Atlas LSH neural networks
With locality sensitive hashing better to forget the h word.
You get thousands of little geometry sensors that 50:50 tell you on which side of a random hyperplane your input data is.
Despite being 50:50 each bit is quite informative.
Several bits together narrow down which geometric region the input is in.
If the inputs have regularities, bits can be codependent while still showing 50:50 on off behavior. That's kind of a subtle point but a gift from random projections - locality sensitive hashing.
Starting from there you can build LSH context dependent neural networks where parameter selection (information routing) is determined using LSH bits.
I've a bunch of notes of say preliminary draft quality.
Maybe start with this one:
https://archive.org/details/atlas-lsh-neural-networks-hierarchical-geometry-rather-than-hierarchical-features
And then click on 'uploaded by" for more should you wish to.
#technology
With locality sensitive hashing better to forget the h word.
You get thousands of little geometry sensors that 50:50 tell you on which side of a random hyperplane your input data is.
Despite being 50:50 each bit is quite informative.
Several bits together narrow down which geometric region the input is in.
If the inputs have regularities, bits can be codependent while still showing 50:50 on off behavior. That's kind of a subtle point but a gift from random projections - locality sensitive hashing.
Starting from there you can build LSH context dependent neural networks where parameter selection (information routing) is determined using LSH bits.
I've a bunch of notes of say preliminary draft quality.
Maybe start with this one:
https://archive.org/details/atlas-lsh-neural-networks-hierarchical-geometry-rather-than-hierarchical-features
And then click on 'uploaded by" for more should you wish to.
#technology
1
| 4,000 mlx | |
| 6 | |
| 1 |