Product Key Memory

Splits large memory matrices into two "half" keys, then multiplies them together to exponentially increase the possible key space, while maintaining efficient search (compute) costs by only performing Top-K operations on each half-key and their combined candidates