"1,000X in Three Years: How Embedded Vision is Transitioning from Exotic to Everyday," a Presentation from the Embedded Vision Alliance
1 like114 views
The document discusses the anticipated 1,000x reduction in cost and power consumption of vision computing over the next three years, driven by deep neural networks (DNNs). It highlights the ongoing improvements in algorithm efficiency, specialized processors, and domain-specific software tools that are making visual perception capabilities more accessible and affordable. As a result, new applications and devices will emerge, enabling more affordable, intelligent, and autonomous systems.
"1,000X in Three Years: How Embedded Vision is Transitioning from Exotic to Everyday," a Presentation from the Embedded Vision Alliance
- 1. Copyright © 2017 Berkeley Design Technology, Inc. 1 Jeff Bier — Founder, Embedded Vision Alliance | President, BDTI May 2017 1000x in Three Years: How Embedded Vision is Transitioning from Exotic to Everyday
- 2. Copyright © 2017 Berkeley Design Technology, Inc. 2 The cost and power consumption of vision computing will decrease by ~1,000x over the next 3 years. 1000x in Three Years Image: BusinessInsider.com
- 3. Copyright © 2017 Berkeley Design Technology, Inc. 3 Visualizing 1000x Image: Daily Mail Image: GalleryHip.com
- 4. Copyright © 2017 Berkeley Design Technology, Inc. 4 • Mobile phones and wireless data went from “inconceivable” to “ubiquitous” • Enabled thousands of new businesses and business models along the way • Example: Facebook has 1.7 billion mobile users, 56% are mobile-only • Vision is poised to be the “new wireless” Why Does This Matter? Image: Wb6nvh.com 1948: First mobile phone 2017 (70 years later): More than 7 billion phones Image: finder.com.au
- 5. Copyright © 2017 Berkeley Design Technology, Inc. 5 Breakthrough detectionaccuracy graph: R. Girshick • Over the past 5 years, deep neural networks have enabled big advances in accuracy for many visual perception tasks
- 6. Copyright © 2017 Berkeley Design Technology, Inc. 6 • DNNs are inherently compute- and memory-intensive • Researchers have mainly pursued accuracy with little regard for compute • XXX Focus on accuracy, not resource use • XXX Massive resource use • XXX But now…. Challenge: DNNs are Very Compute-Hungry Canziani et al., 2017 Digital Trends
- 7. Copyright © 2017 Berkeley Design Technology, Inc. 7 • Deep neural network algorithms are becoming the dominant approach for many types of challenging machine perception tasks • These algorithms replace much more diverse classical approaches • This consolidation has enabled concentration of attention and resources Catalyst: Consolidation Acceleration Cormac Brick, Movidius Szegedy et al., 2015
- 8. Copyright © 2017 Berkeley Design Technology, Inc. 8 How Will We Get There? Image: overclock.net
- 9. Copyright © 2017 Berkeley Design Technology, Inc. 9 How Will We Get There? Image: overclock.net
- 10. Copyright © 2017 Berkeley Design Technology, Inc. 10 1. Algorithms 2. Processors 3. Frameworks, tools, middleware 1000x Efficiency: How Will We Get There?
- 11. Copyright © 2017 Berkeley Design Technology, Inc. 11 • Until recently, most DNN algorithms were developed by researchers seeking to advance the frontiers of machine perception • At top research labs, there’s plenty of compute power available • Now, product developers are focusing on how to deploy these algorithms • Unsurprisingly, they’re finding lots of ways to reduce resource use Improving Algorithm Efficiency Image: Softpedia News
- 12. Copyright © 2017 Berkeley Design Technology, Inc. 12 • Many techniques have been proven effective for slimming down existing algorithms • E.g., reducing data types: • 32-bit floating-point 16-bit float 16-bit fixed 8-bit fixed ? • Pruning • Compression Improving Algorithm Efficiency Tom Michiels, SynopsysHong et al., 2016
- 13. Copyright © 2017 Berkeley Design Technology, Inc. 13 • Even more promising: Design new DNN algorithms from scratch, optimizing both accuracy and resource use • Or, better still: Let tools automatically explore the trade-off space Improving Algorithm Efficiency Minje Park, Intel Samer Hijazi, Cadence
- 14. Copyright © 2017 Berkeley Design Technology, Inc. 14 For decades, chip designers have created specialized processors to get big gains in cost/performance and energy-efficiency Processors Zhang and Brodersen via Sentieys et al. 2014
- 15. Copyright © 2017 Berkeley Design Technology, Inc. 15 Today, dozens of chip and IP core suppliers are creating processors specialized for deep neural networks Processors Pierre Paulin, Synopsys Petronel Bigioi, FotoNation Image: GeekHack
- 16. Copyright © 2017 Berkeley Design Technology, Inc. 16 Software tools translate the intent of the algorithm designer into basic operations performed by the processor • Compilers, function libraries, frameworks, etc. Domain-specific software tools typically create much more efficient code, for a limited range of functionality Software Tools Generate Efficient Code Yair Siegel, CEVA
- 17. Copyright © 2017 Berkeley Design Technology, Inc. 17 Today, dozens of software, chip and IP core suppliers are creating software tools specialized for deep neural networks Domain-Specific Software Tools Venkataramani and Nehemiah, MathWorks Highlights • Automate compilation of MATLAB to CUDA • 14x speedup over Caffe & 3x speedup over TensorFlow
- 18. Copyright © 2017 Berkeley Design Technology, Inc. 18 1. Algorithms 2. Processors 3. Frameworks, tools, middleware 1000x in 3 Years: How Will We Get There?
- 19. Copyright © 2017 Berkeley Design Technology, Inc. 19 • Many types of devices and systems can become safer, more autonomous, easier to use and more insightful • Unheard-of new applications of vision • Many new, visually intelligent devices that are: • Inexpensive • Battery powered • Always on What Does This Mean?
- 20. Copyright © 2017 Berkeley Design Technology, Inc. 20 Today We Have… • OrCam’s interpreter for the visually impaired: $3,500 Image: Daily Mail
- 21. Copyright © 2017 Berkeley Design Technology, Inc. 21 Soon We’ll Have… The “People Person” Tie Tack • Recognizes every person in your network • Face, iris • Whispers their name to you via Bluetooth if you don’t seem to remember! ☺
- 22. Copyright © 2017 Berkeley Design Technology, Inc. 22 Today We Have… • Anki’s “Cozmo” social robot Image: TheVerge.com
- 23. Copyright © 2017 Berkeley Design Technology, Inc. 23 Soon We’ll Have… The Empathetic Teddy Bear • Recognizes your child, his or her friends, your pets, other family members • Can read emotions and interact socially • Sends you cute photos! • (No mobile phone required!) Image: lovethispic.com
- 24. Copyright © 2017 Berkeley Design Technology, Inc. 24 Today We Have… • Production-worthy augmented reality glasses costing $3,700 and weighing 1.3 lbs. Image: Microsoft
- 25. Copyright © 2017 Berkeley Design Technology, Inc. 25 Soon We’ll Have… Image: BrickHouse Security • Production-worthy augmented reality glasses costing $300 and weighing 2 oz.
- 26. Copyright © 2017 Berkeley Design Technology, Inc. 26 Jeff Bier — Founder, Embedded Vision Alliance | President, BDTI May 2017 Thank you!