Edge AI and Vision Insights: April 7, 2021 Edition

LETTER FROM THE EDITOR

Dear Colleague,

It can be hard finding the right people and best technology when you’re creating leading-edge computer vision and edge AI-enabled products… but we have an incredible opportunity for you to connect directly with suppliers and learn about the latest building-block technologies. It’s happening at the virtual 2021 Embedded Vision Summit, and it’s the perfect gateway to propel your current or next project. Plus, there will be 80+ highly relevant, top-quality sessions to gain technical know-how and the latest insights. Check out the expanding list of sponsors and exhibitors to see who you’d like to connect with. Next, learn more about the keynote from UC Berkeley Professor Pieter Abbeel, the general session presentations from Edge Impulse’s Zach Shelby and Qualcomm’s Ziad Asghar, and all of the other exciting presentations and activities at the Summit. And then register today with promo code EARLYBIRDNL21 to receive your 15%-off Early Bird Discount!

Brian Dipert
Editor-In-Chief, Edge AI and Vision Alliance

DEEP LEARNING TRAINING AND INFERENCE ADVANCEMENTS

Reinforcement Learning: a Practical Introduction
In this presentation, Joe Booth, former Vice President of Engineering at Orions Systems (now part of Microsoft, where Booth is a Principal Group Engineering Manager) and an independent researcher, explains how reinforcement learning relates to other machine learning methods, provides examples of real-world deployments, and gives a technical overview of the elements of reinforcement learning. Booth also presents practical advice on when to use reinforcement learning and how to structure problems to use reinforcement learning effectively. Finally, he provides the recommended resources for learning more about this important technique.

Federated Edge Computing System Architectures
With ever-increasing amounts of video and other sensor data, and growing requirements for privacy and low latency, inferencing at the edge is increasingly attractive. But there are many ways to allocate and coordinate computing resources for edge inferencing. For example, to achieve scale and fault tolerance, design principles from cloud computing can be applied to create compute clusters at the edge that are managed by the cloud, an approach called “federated computing.” In this talk, Vaidyanathan Krishnamoorthy, Edge Inference Solutions Architect at Intel, explores a range of edge computing system architectures, with a focus on federated computing. He illustrates how these system architectures utilize Intel CPUs and accelerators to address real-world use cases in retail and industrial applications.

COMPUTER VISION FOR MANUFACTURING LINE INSPECTION

Image-Based Deep Learning for Manufacturing Fault Condition Detection
In this presentation, Jake Lee, Principal Engineer and Head of the Machine Learning Group at Samsung, explores applying deep learning to analyzing manufacturing parameter data to detect fault conditions. The manufacturing parameter data contains multivariate time series sensor signals from a fabrication line. Due to practical manufacturing limitations, datasets are often incomplete, imbalanced and/or not well-formed for deep learning models. To overcome these challenges, Samsung applies new data augmentation methods to train a deep CNN for fault condition classification using deep generative models. Lee also proposes an efficient method to convert multiple time series sensor inputs into a two-dimensional image representation to enable the use of image-based CNNs. Samsung’s experiment results show the fault classification accuracy improvement obtained by applying these techniques.

Deep Learning for Manufacturing Inspection: Case Studies
Deep learning has revolutionized artificial intelligence and has been shown to provide the best solutions to many problems in computer vision, image classification, speech recognition and natural language processing. In this talk, Stephen Se, Senior Research Manager at FLIR Systems (now part of Teledyne Technologies), presents highlights from his experience in deep learning machine vision applications such as manufacturing inspection, defect detection and classification. Deep learning has gained significant attention in the machine vision industry because it does not require the complex algorithm development required for traditional rule-based image processing techniques. Se covers the deep learning workflow from data collection to training and deployment, including transfer learning. Se also presents a few case studies from manufacturing inspection applications.