How is a Fast, Portable, & Accurate Neural Network Possible?

If you develop Computer Vision applications, a good place to be is at the intersection of fast and portable inference—knowing the techniques that make deep neural networks (DNNs) deliver quick, accurate deductions … and ensuring extensibility across a variety of platforms. And choosing a DNN based on design budget, coupled with pretrained-model trial runs using the OpenVINO™ toolkit to validate heterogeneous performance, is a great way to start.

This webinar will give you a jumpstart on all that. Specifically you’ll learn:

  • Basic principles of assessing CNN cost (e.g., MobileNet* vs GoogLeNet* vs VGG*)
  • Algorithmic optimization techniques to improve network performance, including network compression techniques such as network pruning, low precision, and sparsity
  • Effective techniques to improve inference speed and portability, including a comparison of the classic approach vs. a more accurate approach to considering flops, parameters, compute, memory, and heap size
  • How smaller heap size keeps more data closer to compute, runs faster, uses less power, and how the OpenVINO toolkit’s Model Optimizer makes it run ever better
  • How the complexities of pretrained models can be used to create fast and portable new models.

†The OpenVINO toolkit (short for Open Visual Inference & Neural Network Optimization) fast-tracks the development of vision applications from edge to cloud

OpenVINO is a trademark of Intel Corporation or its subsidiaries in the U.S. and/or other countries.

Cormac Brick, Director of Embedded Machine Intelligence, Intel Corporation

Cormac has 20 years of software engineering experience and currently builds new foundational architecture and algorithms for AI and embedded vision to enhance the Myriad VPU product family. Cormac has worked with Movidius since 2007 and has contributed heavily towards the design of the Neural Compute Engine, DSP ISA, deep learning systems software, and classic vision algorithms. Cormac has a B.Eng. in Electronic Engineering from University College Cork, Ireland.

For more complete information about compiler optimizations, see our Optimization Notice.