If you have followed recent news from companies like Baidu, Google, IBM, Microsoft, Qualcomm and Yahoo, you know that deep neural networks (and, in particular, convolutional neural networks) are becoming a very popular means of extracting meaning from images. If you want to understand the practical aspects of using neural networks for vision, you will find a wealth of insights on this topic at the Embedded Vision Summit on May 12 in Santa Clara, California.
First off is the keynote from Dr. Ren Wu, distinguished scientist at Baidu's Institute of Deep Learning. Baidu operates China's largest search engine, and Dr. Wu's talk, “Enabling Ubiquitous Visual Intelligence Through Deep Learning,” will address practical techniques for using convolutional neural networks.
Important topics that Dr. Wu plans to address in his presentation include how to deploy neural networks into challenging applications like mobile and wearable devices with tight cost and power consumption constraints. Also, Dr. Wu will discuss how his team employed clever techniques to create the massive sets of real-world images required to successfully train neural networks, resulting in the recent best-yet published results on the ImageNet object classification benchmark.
Deshanand Singh, director of software engineering at Altera, will present a talk titled "Efficient Implementation of Convolutional Neural Networks using OpenCL on FPGAs." In this presentation, Singh will give a detailed explanation of how CNN algorithms can be expressed in OpenCL and compiled to FPGA hardware.
Another talk in the CNNs-on-FPGAs vein comes from Nagesh Gupta, CEO and founder of Auviz Systems. Entitled "Trade-offs in Implementing Deep Neural Networks on FPGAs," Gupta's presentation will present alternative implementations of 3D convolutions (a core part of CNNs) on FPGAs, and discuss trade-offs among them.
Jeff Gehlhaar, VP of technology at Qualcomm, will present "Deep-Learning-Based Visual Perception in Mobile and Embedded Devices: Opportunities and Challenges." Gehlhaar will explore applications and use cases where developers can benefit from running deep-learning-based visual perception, challenges faced, and techniques to overcome them.
And Bruno Lavigueur, embedded vision project leader at Synopsys, will present "Tailoring Convolutional Neural Networks for Low-Cost, Low-Power Implementation." Lavigueur will share his team’s experience in reducing the complexity of CNN graphs to make the resulting algorithm amenable to low-cost and low-power computing platforms.
In addition to these new neural-network-focused presentations, the Embedded Vision Summit includes numerous other business and technical presentations, along with a second keynote talk from robot vision pioneer Mike Aldred and a technology showcase. The Embedded Vision Summit takes place on May 12, 2015 at the Santa Clara (California) Convention Center. Related half- and full-day workshops will occur on May 11 and 13. Register today, while the limited-time "early bird" discount is still available!
