Amazon Technologies, Inc.

Long Gao

Gang Chen

Matthew James Wood

Eduardo Manuel Calleja

Vinayak Ashutosh Agarwal

Jason Lenox Copeland

Naga Krishna Teja Komma

Ricardo Elizondo Costa

Jiajun Sun

Poorna Chand Srinivas Perumalla

Nagajyothi Nookula

William Shannon Fu

Techniques for machine learning (ML) model adaptation via segment replacement and student-teacher training are described. A model optimizer determines that a ML model sought to be deployed to an edge device includes a source segment that is not supported by the edge device. The model optimizer identifies a replacement segment that is equivalent to the source segment, constructs an adapted ML model by swapping in the replacement segment for the source segment, and trains the adapted ML model, optionally from the source ML model using a student-teacher training procedure. The trained adapted ML model is then deployed to the edge device, where it can be successfully run with minimal (if any) degradation of performance compared to the original source ML model.

Machine learning model adaptation via segment replacement and student-teacher training

Techniques for training a machine learning model based on captured images are described. A method described include filtering a first set of collected images using one or more machine learning models; labeling the first set of filtered, collected images using a data labeling service using a service of the provider network; training a machine learning model from a machine learning algorithm using the first set of filtered, collected images using a service of the provider network; and causing deployment of the trained machine learning model onto a device.

Determining data relevance and labeling, model updating, and deployment

Techniques for high-performance machine learning (ML) inference in heterogenous edge devices are described. A ML model trained using a variety of different frameworks is translated into a common format that is runnable by inferences engines of edge devices. The translated model is optimized in hardware-agnostic and/or hardware-specific ways to improve inference performance, and the optimized model is sent to the edge devices. The inference engine for any edge device can be accessed by a customer application using a same defined API, regardless of the hardware characteristics of the edge device or the original format of the ML model.

High performance machine learning inference framework for edge devices

High perforamance machine learning inference framework for edge devices

Systems, methods, and apparatuses for determining data relevance and labeling, model updating, and deployment

Techniques for machine learning (ML) model knowledge distillation and automatic retraining are described. A model adaptation controller obtains samples generated by an edge device and inference values generated based on the samples by a deployed ML model of the edge device. The model adaptation controller runs inference on the samples using a different ML model to generate inferences that can be used to determine whether the performance of the deployed ML model is lacking. If so, the model adaptation controller can retrain the deployed ML model using samples with ground truth values generated by the different ML model, resulting in a light-weight retrained model that can be provisioned to the edge device. This retraining process may be performed iteratively to automatically improve and adapt the ML model running at the edge device.

Knowledge distillation and automatic model retraining via edge device sample collection

Growing community of inventors

San Jose, CA, United States of America

Long Gao