Researchers propose amortized MIPS using learned support functions for neural prediction.

Amortized MIPS trains neural networks to directly predict maximum inner product search solutions, treating the MIPS value function as the support function of the key set to amortize repeated solving costs.