Deep-learned asymptotically optimal matrix product using approximate bilinear algorithms
| Main Author: | Shrohan Mohapatra |
|---|---|
| Format: | info publication-preprint Journal |
| Terbitan: |
, 2020
|
| Online Access: |
https://zenodo.org/record/3951841 |
Daftar Isi:
- This article introduces a novel neural network framework for the approximate bilinear algorithm that scales asymptotically with the input size. The network actually learns from the deviations from the expected form of the identity and in the progress of learning the matrix multiplication exponent gradually decreases via a probabilistic trail. The probability density function associated with the learning algorithm decides the time the algorithm takes to converge and arrive at an appreciable form of the identity. Consequently, one can also accelerate commercially accepted context-free parsing by the reduction to Boolean matrix multiplication and then using the proposed scheme.