GTC 2020: Automating DNN Design for DRIVE AGX: Platform-Aware Neural Architecture Search

GTC 2020 S21666
Presenters: Arash Vahdat,NVIDIA; Le An,NVIDIA
Abstract
In the past few years, wide applications of deep neural networks (DNN) have contributed to significant progress in various fields such as image classification, object detection, and segmentation. Most of the successful DNNs, such as VGG and ResNet, are designed by humans, which requires in-depth domain expertise and effort. While DNNs have become deeper and wider, the need for fast inference is increasing on (edge) computing devices, while accuracy must be maintained. Therefore, developing state-of-the-art neural networks for resource-constrained applications has become challenging. We’ll present our progress on the automated design of neural networks using hardware-aware neural architecture search (NAS) techniques. We show concrete end-to-end examples from differentiable and latency-reflected search of optimal network architectures to their deployment on NVIDIA’s DRIVE AGX platforms using TensorRT for autonomous-driving-related applications.

Watch this session
Join in the conversation below.