Amazon Technologies, Inc.

Gm Cruise Holdings LLC

Shuai Zheng

Sachithra Hemachandra

Yida Wang

Mu Li

Trishul Amit Madhukar Chilimbi

Yue Zhao

George Karypis

Li Zhang

Zheng Zhang

Sanjiv Ranjan Das

Zhen Zhang

Sheng Zha

Navaneeth Bodla

Justin Chiu

Xiao Zhang

Gia Tri Nguyen

Ajaya Hanasoge Shankaranarayana Rao

Zhijiang He

Xin Huang

Shenghua Yue

Amirreza Shaban

Machine learning models for controlling AV operations may be evaluated based on sliced data. A feature category critical for an environmental condition may be identified. The environmental condition is a condition in an environment where an AV performs an operation. A sub-dataset, which comprises sensor data capturing the environmental feature dataset, may be extracted from a dataset that comprises sensor data capturing the environment. The sub-dataset may be input into machine learning models that can classify the feature category. Performances of the machine learning models can be evaluated based on their outputs, which are generated based on the sub-dataset, and a ground-truth classification of the environmental feature. The output of each machine learning model comprises a classification of the environmental feature. A machine learning model may be selected based on the evaluated performances of the machine learning models and may be used to control driverless operations of AVs.

Using sliced data to evaluate machine learning models used for autonomous vehicle operation

Autonomous vehicles utilize perception and understanding of road users to predict behaviors of the road users, and to plan a trajectory for the vehicle. Understanding subtypes and attributes of vehicles may help autonomous vehicles better predict behaviors of and react to vehicles in a variety of road situations. To offer additional understanding capabilities, an additional understanding model is added to the perception and understanding pipeline to improve classification of vehicles and extraction of attributes of the vehicles. The exemplary architectures of the understanding model balance recall and precision performance metrics and computational complexity.

Perception and understanding of vehicles

A resource set which includes multiple servers with a respective plurality of training computing devices is identified for training a machine learning model. The resource set is subdivided into partition groups, such that each partition group can store a respective replica of state information of the model. The model is trained using the partition groups. The training comprises a multi-stage gathering of a portion of the state information at training computing devices of a particular partition group. Different types of communication channels between training computing devices are used in respective stages of the gathering, including inter-server communication channels in one stage and an intra-server communication channel during another stage. A trained version of the model is stored.

Distributed training of machine learning models

Techniques for machine learning (ML) and natural language processing (NLP) are described. One technique enables the creation of a clean training dataset through just a few API calls. Another technique provides an automated process for generating a domain-specific lexicon, which is then used to generate ML training datasets, in a manner that requires little to no human labor. Another technique gathers ML training data from domain-specific public sources, which are more likely than typical public sources to contain focused terminology and to be free from errors, thus resulting in trained ML models that provide more accurate inferences.

Domain-specific language models

Growing community of inventors

Santa Clara, CA, United States of America

Shuai Zheng