[closetCS]

Its pretty smart to use neural networks to exploit the non uniformity of the data, but I thought they were going to have replaced just the hash function part of a bloom filter, instead of the entire algorithm. I wonder if it could preserve the speed of the bloom filter while exploiting potentially non uniform data.

[stochastic_monk]

There's probably something worth using in this tradeoff space, though this isn't it.

[sitkack]

Maybe the network can be used to construct the hash function after it has learned the shape of the data.