Superbloom: Bloom filter meets Transformer
This provides an alternative method for problems with large vocabulary size, which is incremental as it adapts existing techniques like Bloom filters and Transformers to a new context.
The paper tackles the problem of handling large vocabulary sizes in machine learning by extending word pieces to opaque ids using Bloom filter hashing and a multi-layer Transformer, achieving high accuracy that outperforms similar-sized models without hashing and larger models with sampled softmax under the same computational budget.
We extend the idea of word pieces in natural language models to machine learning tasks on opaque ids. This is achieved by applying hash functions to map each id to multiple hash tokens in a much smaller space, similarly to a Bloom filter. We show that by applying a multi-layer Transformer to these Bloom filter digests, we are able to obtain models with high accuracy. They outperform models of a similar size without hashing and, to a large degree, models of a much larger size trained using sampled softmax with the same computational budget. Our key observation is that it is important to use a multi-layer Transformer for Bloom filter digests to remove ambiguity in the hashed input. We believe this provides an alternative method to solving problems with large vocabulary size.