Innovation Panel Discussion at TechCon Socal 2020: How AI is impacting our future

Yeah, I think just to echo Arjun's comments, there's this convergence happening towards ubiquitous sensing plus the compute storage and drop in cost, and that's really hitting every industry. Industrial is kind of farther ahead when it comes to sensing, so they've invested— the typical compressor might have 400 sensors I, on a good day, have one heart rate variability sensor. So I remember we used to think it'd be great if we were as valuable as these compressors, but they've really invested heavily to add the sensors and have these AI-powered engines monitoring their health predictively. Again, something I think is the future of human healthcare. It'll be great when we get more of that online sensing. Yeah, I think that the next opportunities I'm excited about— I agree, there's going to be continued innovation in core algorithms and deep learning. You know, we have Convolutional neural networks, and then on NLP recently found out transformer models were fundamental to breaking the SuperGLUE benchmark. And we're going to continue to see fundamental algorithmic advances, but separate from that, more like composition of what— there's a lot of really narrow niche, you know, AIs out there right now, but putting together in a good user experience I think is going to be good. I think, you know, the massive amount of data that's available online now with text, so NLP is going to be a key advance. And systems that monitor humans and then improve our interaction, there's no question in my mind that that will improve our interface. Like the affective computing, right, with emotion being an early kind of pioneer in that. I think we'll see more of that, whether it's— there's this term centaur, which is the human-AI hybrid. And in industrial, you have a lot of people operating machines that could blow up or kill everyone on the entire platform. And so There's a key need to make sure, is that person tired or fatigued or distracted or processing some emotional thing? And it's all those kinds of human factors. And so better systems environmentally to track that. I think we're getting more and more sensors for humans, which will be good for improving

that. Ken also has some comments on that, so if possible.

I'll be brief, but I want to give a sort of a very practical example of what folks were just talking about. I really am a strong believer in the benefits of augmenting humans, and I think that that's going to be something that we'll see. And if you think about how sort of commercial airlines happen today, for the most part, pilots, you know, they're there for the takeoff and they're there for the landing. Pretty much everything else in between is is driven by computers. Boeing just found out they're a software company recently. And so if you take that analogy to places like surgery, I think you'll begin to see— and it's going to come from a combination of AI, robotics, natural language processing, where you see immersive technologies. We have a company that I'm an investor in. It's ProprioVision up in Seattle. And basically, it's using light field technology To create an immersive environment for surgeons. And it comes from a use case where this pediatric neurosurgeon who says, I'm wearing these very heavy loops, the sort of microscopes that they attach to their glasses, and they're really heavy and they're not very ergonomic, and those are long surgeries in a very confined space if you think about babies or children that they may be performing on. And it takes really, really fine motor capability to do this kind of work. Well, and today it's a 1D microscope, you know, but if you have an immersive environment, you can, you know, go from 0 to 50, 0 to 100 magnification. You can see that vein on the other side of the organ. You can see the coloration. It's just a different world. And then if you have augmented reality, which delivers just the right information that you need at exactly the right moment, As opposed to having to step out, do research, then scrub back in. I mean, and then from a robotic standpoint, the ability— once you digitize the field of view, you can actually place that screw in the spine or at exactly the right angle that you're shooting for, at exactly the right penetration, 0.12 millimeters. And so does this diminish the doctors, the surgeons? No, the discernment, the knowledge that the surgeon brings, the decision-making, that's all still needed. But you've actually sort of democratized surgery because the best surgeons in the world are the best surgeons in the world because they get the most time on target. Okay, but if you can stream this in real time and other surgeons can watch from multiple angles this surgery, they come up to speed faster. So there's enormous benefits of the technology to augmenting humans, and it helps society. So I just want to give you a real practical example of where some of this is going, and it's not 10 years away. This is like 2, 3 years away, really.

Maybe I'll pose a question to Jie Lei. Jie Lei, since talking about some of this, some of the AI in settings such as medical fields or health fields, You talked about some of the autonomous car use cases. There's certainly a confluence of different technologies coming together, right? 5G connectivity, you have AI, much more sophisticated at the edge, right? How do you kind of see that playing out specifically from your vantage point? Because you are looking at 5 years out, maybe more than that from an R&D perspective. So your focus is, in many cases, like a lot of the comments you're hearing, it's probably going to be somewhat nearer term. You're looking way ahead. So what does your longer-term R&D lens tell you about further out? What are the use cases that will get solved, maybe 5-plus-year time horizon?

Thank you. Thank you for the opportunity. I think I do have some thoughts from a research front. I think Ken just mentioned about deep tech perspective. How do we think about from a longer time horizon? I think not just Qualcomm. I think the whole machine learning research community can aim high and aim more ambitiously. I would say number one, actually, as much as deep learning has shown a lot of promise, but I have to say it does carry a lot of limitations. So one of the primary sort of limitations I felt tying to the data cleansing and all the other in terms of constraint actually is about generalization of model. So when we talk about generalization of model, meaning when you train a model for a certain dataset, can it be very easy or convenient being transferred or leveraged to a new field, new context, and new circumstances? But I'd say, obviously there's a lot of cool research in that space, but it's still very, very early stage in terms of how to make sure model more generalization can work for multitask learning, for continuous learning. I think there's a tremendous need and a demand to make sure AI can be much more broadly usable across different spaces. And the second thing I'd also say, this is one area actually Qualcomm is joining a lot more into the research community now. Is a lot of people think about, hey, if I just simply throwing a deep neural net, it simply can just work because I can do end-to-end training. It sounds magic, right? It's really cool. But on the other hand, from the opposite side, I think as Qualcomm, as a leading signal processing vendor from modem, from a multimedia perspective, we also realize through a lot of the practice we have found, if you can combine the design knowledge from human, which is coming from your domain expertise, With machine learning and the deep learning training capability over trying to identify deeper unidentified features from all the data itself, we're finding out that gives you 2 tremendous benefits. It gives you data efficiency. It also gives you the generalization benefits. In addition to this, I think a couple of things that I probably mentioned very quickly. One is also people talk about, hey, deep learning largely is only doing pattern recognition. How do you take that step forward? How you can do reasoning based on causality where you can reason over symbolic structures? I think this is a tremendous amount of needs for AI to take the next steps. And also if you think about privacy and how people are looking at using federal learning technique, keep your data on the device but you can doing a training over crowdsourcing over a large amount of device. There's also a lot more research promise in that space. The last, not least, from going back to the comment I think we made early on, from the power efficiency perspective to driving the model to be the most convenient and powerful to be deployed on device on a data center, in energy consumption, there's still tremendous need and demand to driving you can looking at the model from a use case development to compression and down to compiler to the hardware design. end-to-end design. I think there's a lot of need to— we should continue to do a lot of innovation work in that space.

Yeah, okay. Augustin, maybe a question to you. There are a lot of companies, and I think Ken sort of hit upon that, which is there's a lot of misses in the AI. And there's 2 things. There's one, you know, the sort of disillusionment or disenchantment even with the large companies, number one. Number 2 is many times they just have a number of other initiatives that are going on. So how do you really start to balance it out and say, look, I I did some experiment, maybe I put together a team, didn't work out. How do I go back and set up a process where I can—

Ken's exactly right. There are a lot of companies that have invested a lot of money in this stuff and have not seen a whole lot of return. I've talked to plenty of those kinds of companies.

I think the biggest mistake that companies, medium, even large-sized companies make, is they overinvest in products and underinvest in people and processes. Because it's really easy to sign a contract, get a data warehouse, In the door, and it's like, okay, I'm doing something about the problem, right? And then they turn around and they hire like a junior engineer to sort of somehow transform that data into actual decision-making insights that they can, they can actually use.

I call it the Underpants Gnomes problem. Who's familiar with South Park here? Yes. Okay, so the Underpants Gnomes are a secret society of gnomes. Their plan for world domination is step 1, grab the underpants. Step 2 they're not clear on, and step 3 is profit. So companies like just swap out the underpants with data, and I feel like that's exactly the same decision-making process. And the problem is it's exactly backwards. You have to start with step 3, like what is the decision that I need to improve? What is the, what is the dollar value of the thing that I'm trying to get better at? And then you work backwards. Okay, well, what systems do I need to support that better decision-making? Well, what data do I need to support those systems? When you look at it that way, then getting back to Ken's point, like, well, now you can actually think about things in a way that makes business sense. And that's the thing that a lot of companies still to this day have a lot of trouble with.

Can we get some more questions maybe from the audience? Yes, please.

General question on sort of not from what can be done with current methods and algorithms, but the problems that we actually need to solve. A few days ago, Microsoft announced very aggressive climate goals, and I think a lot of other big tech companies will follow because there's a lot of pressure from the society and everything. So how does that create all these clean tech and, you know, focus on energy efficiency, maybe focus on reducing emissions? How does that affect— mainly question mostly for Jia because you're working in the power efficiency, power optimization. Does that create new opportunities? Does that make kind of more important to focus not just on the overall cost of deploying an AI system or building a more effective chip, but specifically for the emissions portion?

For the emission from— Just from energy

consumption, for example.

General energy consumption, yeah. So coming back to energy consumption, actually, people probably don't realize— I don't have the exact number— if you're training AlphaGo model, let's say, you know, DeepMind, right? It's probably going to take one car to driving, I think, on the road maybe for at least a few months in terms of how much the CO2 has been generated. So I think enormous amount of things I think people have to deal with in that sense. But coming back to from a lot of what we do on our end, you know, very disruptively we're looking at probably we can share some knowledge here. We fundamentally looking at the hardware architecture, how we can break through the barrier. We can take another 10x down in terms of energy consumption. So the way we definitely see going to take a few things through hardware architectural revolution itself, and also in terms of hardware with the memory co-design. For example, we have talked about computing memory and how we choosing the best memory type. To help you fit to driving the ultra-low power consumption from a just silicon perspective. But on the other hand, I think Ying said from another panel, I think it's tremendously important to think about the software-hardware co-design. So meaning if you can looking from a use case to model architecture design to compiler to the end chipset, if you can just lowering down the graph, what we call interrepresentation, just automatically directly to the hardware in a whole end-to-end manner. I think this is an early concept. There's early proof of concept, but make it really to be mature to driving the whole industry, I think still a lot more can be done. Thank you.

We'll have time for one more question. Please go ahead.

My question is about privacy, but it's not meant to be a political question. It's meant to be a practical question. When I look at the last 20 years, there's been privacy concerns along the way, you know, the introduction of cameras and phones and social networks. Yet as a culture, the culture has always evolved to accept these privacy trade-offs. And so my question is, as we deploy more and more AI over the next 10, 20 years, do you think the status quo of culture gradually accepting privacy trade-offs will continue, or do you think there's a cliff ahead or a mountain ahead where the I think it

already is. If you look in the ad space, everybody's obsessed with data clean rooms. Like, you have to sort of go into this very special place to look at ad targeting data. So, I mean, it's definitely already happening. In terms of like a far-out tech, I would say homomorphic encryption, if you could do that. If you could sort of do operations on data without actually seeing the data, it's all encrypted sort of end to end. I mean, if we could figure that out, that'd be pretty cool. I think right now there's— I think it's moving more towards the fear kind of response. And there's that documentary, The Big Hack, or Great Hack, on Netflix about Cambridge Analytica. And the more data you have out there, maybe there's risk, A, for manipulation, and B, for solving cancer, curing mental health, and making our lives better. And it's a trade-off, right? So I think it's— an interesting discussion. I think there's a shift towards more companies focusing on positive advertising-driven business models. I call it— it's not the attention economy, it's the addiction economy. And I'm looking forward to a shift away from that, hopefully driven by Southern California.

So one more comment, I think, going back to the research side. How to combine the mathematical power from statistical processing and deep learning together where you can rely not just simply on the traditional security approach but still in a deep learning wonderland. We can say, hey, we can anonymize the data and provide a great privacy guarantee to the people. I think there's a lot of cool research going on in that space and coupling with how the deep learning model can be trained and deployed in a more flexible way, there's a different topology sort of domain you can play coupling with advanced deep learning methods. which can provide further advancement from the privacy angle too.

Comment from Ken? I'll just make a— in general, my view is that as a society, as you rightly point out, we tend to get bought off. So that, you know, to the extent that we get more benefits offered to us, we're willing— and it's a broad generalization— we're willing to give up some of our privacy for the goodies. And businesses know that. They navigate it. People that wanted protection from particularly advertising, you know, there's a way to be protected. But if you look at the data from GDPR and even in California, early data, very few people are opting out or going to the steps that, you know, they need to take to be forgotten. So it's an emotional time. Nobody's individual information is actually worth that much. Anonymized aggregate data for groups and subgroups, worth a lot. So over time, I think this is an issue where people like big companies, maybe the big A, they'll say, gosh, if you opt in to allowing us to use on an anonymized aggregate basis the data that we're collecting, I'm sure you'll get a free phone. Phone every 2 years or whatever. And if you look at the data, younger folks, my daughter, they're just not that into it, into the privacy side. So it's a broad generalization, but over time I think we get bought off.

All right. I think we're going to end it here because we're out of time. But Alexa is going to know more about us than it ever did 6 months from now. So that's all happening. I want to thank all the panel members. for, for their good contributions. Thank you.