Learning Ultra-Lean Classifiers with Structured Functions and Binary Lifting


At a glance

Stella Yu 

model compression, manifold learning, complex-valued deep learning, binary lifting

Deploying classifiers optimally learned offline from big data requires model compression and quantization, often with lossy performance and yet not enough compression. Our idea is to train a lean structured model to begin with, by building in the type of invariance the data/task desires, the type of layer functions that enforces sharing, and by delivering a binary representation that can be executed using logic-only operations. An ultra-lean classifier can thus be efficiently learned from the start to the end and deployed with far fewer MACs than a conventional deep classifier network.