Machine Learning for the Real World: What is Acceptable Accuracy, and How Can You Achieve It?


Date:  Tuesday, September 15, 2020

Start Time: 1:30 pm

End Time: 2:00 pm

The benefits of running machine learning at the edge are widely accepted, and today’s low-power edge devices are already showing great potential to run ML. But what constitutes acceptable accuracy when applied to real-world, real-time use cases? In this talk we explore what constitutes acceptable detection accuracy for specific use cases, and how this can be measured. Looking at which ML models are meeting the challenges and which fall short, we focus on how techniques like transfer learning can help fill the gaps when weaknesses in detection accuracy are found.

Track

Session Speakers

  • Tim has worked at the cutting edge of embedded vision and machine learning enablement for the past eight years at Arm, from helping partners understand how to target compute onto then-emerging mobile GPUs, to initiating and driving requirements for the open-source Arm Compute Library and Arm NN framework. Now part of Arm’s strategy group, Tim runs the product team within the Insight incubation unit, designing new forms of edge-to-cloud distributed machine learning verticals covering smart buildings, smart cities, industrial health and safety, healthcare and more. Tim is a long-time attendee, contributor and advocate of the Embedded Vision Summit, remembering the event in its early days, long before the move to long trousers. Previously to Arm, Tim ran a research and development team at the BBC, exploring multimedia techniques and technologies, and served for a period as an entrepreneur building a number of technology start-ups.


Share

Machine Learning for the Real World: What is Acceptable Accuracy, and How Can You Achieve It?