Abstract: A resource management system is disclosed herein that quickly and dynamically tailors application resource provisioning to real-time application resource consumption. The resource management system may service application requests using resources selected from a pool of servers, the pool of servers including a mixture of virtual server resources and serverless instance resources. The serverless instance resource may comprise software objects programmed using a machine image reflecting one or more states of a virtual application server booted using application-specific program code. Supporting an application using serverless instances enables dynamic scaling of application resources to support real-time application servicing loads.

Abstract: Systems and methods for training and using machine learning models to classify vehicles from highway radar systems are provided. The training systems may use auxiliary radar processing to separate events by lane, length, and/or speed, and then use separate event data groups pooled from similar or proximate lanes, lengths, and/or speeds to train multiple models. At estimation time, incoming events may be grouped using similar groupings as those used during training to select which model to use. An incoming event may be applied to the neural network operations of the selected model to generate an estimate. Generating an estimate may involve successive applications of multiple linear convolutions and other steps along varying or alternating dimensions of the in-process data.

Abstract: Systems and methods for training and using machine learning models to classify vehicles from highway radar systems are provided. The training systems may use auxiliary radar processing to separate events by lane, length, and/or speed, and then use separate event data groups pooled from similar or proximate lanes, lengths, and/or speeds to train multiple models. At estimation time, incoming events may be grouped using similar groupings as those used during training to select which model to use. An incoming event may be applied to the neural network operations of the selected model to generate an estimate. Generating an estimate may involve successive applications of multiple linear convolutions and other steps along varying or alternating dimensions of the in-process data.