Amazon Technologies, Inc.

Michele Pratusevich

Jeremy Samuel De Bonet

A system that can create an approximated model from a trained machine learning model (such as a neural network) where the approximated model can operate using fewer computing resources than the original trained model. The system can create the approximated model without the voluminous training data used to create the original trained model. The system can rely on just the data describing the trained model and an indicator as to how closely the approximated model should correspond to the original model and/or the desired savings of computing resources. Various lossless and/or lossy approximations may be performed to obtain multiple approximated models that may be substituted for the trained model during runtime operations to achieve significant speed/cost savings over operation of the original trained model.

Growing community of inventors

Somerville, MA, United States of America

Michele Pratusevich