The Cutting Edge of AI — Part 1
Notes from the day one of Exponential View talks and panels at CogX Festival, 11–12 June, London
Exponential View ran one of the five main stages at this year’s CogX festival in London. We gathered some of the world-leading researchers, investors, founders and scientists to discuss the cutting edge of AI, compute, deep tech and human transformation — many of them are EV readers (note: if you’re not on the list yet, subscribe here).
I’ll aim to share some of the key points from a number of talks during the two days of the conference, this post being dedicated to Day 1. For full agenda and videos for each talk, see the very end of this post or this YT playlist.
Notes on the future of hardware & compute
Large part of the first day on the Cutting Edge stage was dedicated to hardware and architectures for the AI age.
ArkInvest’s analyst, James Wang, spoke about the future of AI hardware.
- There’s three axes of improvement that we can use (and have used since the dawn of silicon chips) to improve the performance we need. One, transistor scaling, which is another name for Moore’s Law. Moore’s Law is not dead, it just got much more expensive. Two, improved chip architecture, laying transistors in different ways. Three, deploying more servers, as Google did with its first AI chip between 2016 and 2017 — chips didn’t get a lot smarter in that time, the transistors didn’t shrink substantially, but the number of servers the company was putting out became monstrous.
- How do we close the gap between the performance we need and performance we get? Primarily, in the near term, we need much better chips. Moore’s Law has another 10–15 years left, but we can’t depend on it. TPU-based designs are better architectures, but not sufficient. Software deserves more attention, as there’s around 100x performance just in software. More efficient algorithms the most promising: doing one-shot learning which reduces the number of samples we need; today we need 1000 examples, but if we shrink it to one, that gives us 1000x improvement. New ways of training ResNets such as superconvergence are promising, as well. New substrates, such as optical, analog and quantum are needed to take us beyond the Moore’s Law. If all else fails, the existent system will keep us going, by scaling up (increasing chips in a server), we can scale out.
Simon Knowles, founder of Graphcore, spoke about the most complex logic chip ever made, Graphcore’s Intelligence Processing Unit (IPU).
- Machine intelligence is the greatest driver to re-invent computing since the invention of computers, 75 years ago. This is the first time we’ll reinvent compute from scratch.
- Requirements from compute are only set to grow. Customers demand a million to infinite more compute power from companies like Graphcore, and we’ll see in the coming years many of the devices that weren’t computers in the past, become computers.
- While Moore’s Law will feed another 3x in power, if we want to get 1000x, we must employ new computing architectures and connect many chips together. IPU design drivers will enable the new world for the development of software and resolve the question: how do you get a million processors to talk to each other and remain in order? Bulk synchronous parallel enables the machine to alter between the phase of compute (processors execute data only in local memory and don’t communicate with each other), and exchange phase (they all exchange information ready for the next compute phase, memory-to-memory data movement).
- The roofline model — where all communications in computing operate at the same time maximally — is broken. IPUs distribute lots of very small tiles of memory and processor all over the chip—with 23bn transistors and the ability to execute 7k programs in parallel, this is the most complex logic chip today. Ultimately, we should bulk many of these chips together.
Notes on data and edgeification
Ian Hogarth, the founder of SongKick led a panel discussion dedicated to data and edgeification, joined by Pete Warden, Technical Lead at Tensorflow’s Mobile and Embedded Team, Jon Crowcroft, Marconi Professor of Communications Systems at the University of Cambridge, and Hugo Pinto, Managing Direct at Accenture Digital.
- Some of the underlying trends that motivate machine on the edge are latency, inference in high performing situations, more federated learning, privacy issues. Large companies are interested in short term benefits and smart infrastructure. For companies like Google, interactivity makes a killer application and drives user satisfaction (think Maps or Gmail).
- The impact of edge on businesses is not in delivering insights. The impact is delivering insights in front of the users for a particular asset they’re assessing in real time.
- Metcalfe’s law — internet is super linear — the value of telecommunications network is proportional to the square of the number of connected users of the system. When you centralize things in a cloud, you silo every surface in that cloud service and have sub-linear growth because of the costs associated with it. The increase in value with each additional device will bring back the value of each participant in the system.
- Companies will have brand new value propositions to incentivize edge use cases. Once someone realizes they don’t have to build all the infrastructure, but redesign what they’re doing on the edge, we’ll see a boom in the number of new startups.
- Explosion of specialized apps. DSPs and microcontrollers already have good support to run neural networks. Need to push for further connection of microcontroller engineers with neural network engineers.
- Challenges: keeping stand-alone object simple enough and do enough QA that they’re like other devices we put out in the world like washing machines and windscreen wipers that we tested and understand thoroughly enough that we don’t have to sweat the security process. At the moment, computer programs are narrow at what they’re looking at. If we could have a neural network that’s aware of the context, we could increase device security. Having a range of behaviors contributes to resilience, need a larger gene pool of behaviors.
Notes on the future of urban shared autonomous transportation
Stan Boland, founder of Five.AI, spoke about his vision for mobility, and how the world can gain from European AV development.
- 160m Europeans live in towns and cities with more than 100k people in them. They spend 2.5k dollars/year on travel, which makes urban travel a $400bn annual market in the next 10 years. Eight hundred billion passenger kilometers per year in the UK are dominated by car, van, and taxi.
- Shared urban transport is the solution to an increasingly congested, slow and expensive transportation. The existence of heterogenous and autonomous fleet that would support public infrastructure would bring the cost of urban transport from $18 per day/person to $6 per day/person.
- Technical challenges: perception and prediction. Urban driving is hard: requires high precision and high recall at the same time; the number of things we want to classify is higher, and the number of actors we care about goes from 4–5 to 20. With the current state of vision and AI, we’re still going to miss 30% of cyclists. But massive improvements are happening. FiveAI particularly relies on 3D object detection training to by building full 3D simulations of the real world, creating digital twins of real cities, being able to change the weather, light, object behavior, colors, etc. We also need to make use of as much technology that exists as possible, we can’t rely on any one individual sensor.
- Vast majority of the companies in the space are US and Chinese. Europe is behind the curve. However the talent exists, market exists, the job is to bring these together. Doing so would be beneficial for autonomous driving elsewhere, because European streets are more complex and harder to train on.
Agenda and videos
Danny Lange, VP of AI/ML, Unity — Machine Learning as an AI Toolkit (video)
Adrian Weller, Programme Director, The Alan Turing Institute — Frontiers of AI Research (video)
Panel “Beyond Algorithms” hosted by Libby Kinsey, Head of Technology and Lead Technologist for AI and ML, Digital Catapult Beyond Algorithms, with Jack Clark, Strategy and Communications Director, OpenAI, Danny Lange, VP of AI/ML, Unity, Adrian Weller, Programme Director, The Alan Turing Institute (video)
Casimir Wierzynski, Senior Director, AI Product Group, Intel — Trends in AI Systems (video)
Simon Knowles, Co-Founder and CTO, Graphcore — Accelerating Next Gen Machine Intelligence with IPUs (video)
Igor Carron, CEO, LightOn — Optical Computing Leading the AI Scale-Up (video)
James Wang, Analyst, ArkInvest — Future of Hardware for AI (video)
Panel “Future of Hardware” hosted by Azeem Azhar, Founder of Exponential View, with Casimir Wierzynski, Senior Director, AI Product Group, Intel, Simon Knowles, Co-Founder and CTO, Graphcore, James Wang, Analyst, ArkInvest, Igor Carron, CEO, LightOn (video)
Marta Garnelo, Research Scientist, DeepMind — Symbolic Methods are Coming Back (video)
Jack Clark, Strategy and Communications Director, OpenAIThe Edgeification of Intelligence — Defining the Cutting Edge of AI: Where We Are and Where We’re Going (video)
Joe Baguley, VP and CTO for EMEA, VMwareMachine — The Edgeification of Intelligence (video)
Pete Warden, Technical Lead, Tensorflow’s Mobile and Embedded Team — Learning on Tiny, Cheap Devices (video)
Jon Crowcroft, Marconi Professor of Communications Systems, University of Cambridge — Confidential Cloud and Edge Computing (video)
Panel “Data & Edgeification of Intelligence” hosted by Ian Hogarth, Co-Founder, SongKick with Pete Warden, Technical Lead,Tensorflow’s Mobile and Embedded Team, Jon Crowcroft, Marconi Professor of Communications Systems, University of Cambridge, and Hugo Pinto, Managing Direct of Accenture Digital (video)
Stan Boland, CEO, Five.AI — The Journey to Safe Urban Autonomy (video)
Miles Brundage, Research Fellow, University of Oxford Future of Humanity Institute — Understanding AI Progress (video)
Libby Kinsey, Head of Technology and Lead Technologist for AI and ML, Digital Catapult — AI Plumbing: Bringing Cutting Edge AI Products and Services to Market (video)
Dr Jeni Tennison, CEO, Open Data Institute in conversation with Carina Namih, Parter Episode1 Ventures — What Should Our Data Rights Look Like? (video)
Azeem Azhar’s Exponential View is the #1 newsletter about exponential technologies and their impact on the society. Subscribe → www.exvw.co
