How AI is powering the next generation of security systems with Deep Sentinel's Dave Selinger

Learn why traditional home security systems tend to fail and how Dave’s love of tinkering and deep learning are helping him and the team at Deep Sentinel avoid those same pitfalls. He also discusses the importance of combatting racial bias by designing race-agnostic systems and what their approach is to solving that problem.
Angelica Pan

Listen on these platforms

Apple Podcasts Spotify Google YouTube Soundcloud

Guest bio

Dave Selinger is the co-founder and CEO of Deep Sentinel, an intelligent crime prediction and prevention system that stops crime before it happens using deep learning vision techniques. Prior to founding Deep Sentinel, Dave co-founded RichRelevance, an AI recommendation company.
Learn more about Deep Sentinel
Follow Dave on Twitter
Join his machine learning Meetup

Show notes:

Topics covered:

0:00 Sneak peek, smart vs dumb cameras, intro
0:59 What is Deep Sentinel, how does it work?
6:00 Hardware, edge devices
10:40 OpenCV Fork, tinkering
16:18 ML Meetup, Climbing the AI research ladder
20:36 Challenge of Safety critical applications
27:03 New models, re-training, exhibitionists and voyeurs
31:17 How do you prove your cameras are better?
34:24 Angel investing in AI companies
38:00 Social responsibility with data
43:33 Combatting bias with data systems
52:22 Biggest bottlenecks in production

Transcript:

Dave:
We have this 7-Eleven that gets burglarized literally once a week. Guy walks up with a crowbar and he swings at the door, and that's the end of our video because our guards get on and say, "Hey, jerk, get out of here. The police are on their way," and the guy walks away. Whereas if you had a dumb camera, you get this really cool video that for the next 45 seconds, you see this guy banging on a window. So what we've had to do is we've had to train the market that like, hey, prevention is possible.
Lukas:
You're listening to Gradient Dissent, a show where we learn about making machine learning models work in the real world. I'm your host Lukas Biewald. Dave Selinger is the co-founder and CEO of Deep Sentinel, an intelligent crime prediction and prevention system that stops crime before it happens using deep learning vision techniques. Prior to founding Deep Sentinel, Dave co-founded RichRelevance, an AI recommendation company. Super excited to talk to him. Maybe you could actually start by just describing what your company does and how it uses deep learning.
Dave:
Sure. So Deep Sentinel is a physical security company. So we have cameras and we actually protect facilities. It's not cybersecurity protecting computers. We are kind of a competitor to ADT, which I'm sure you've seen the lawn signs for and the whole premise behind that business was, hey, ADT, doesn't work. A lot of people don't know this, but police departments across the country, basically don't respond to burglar alarms because they're 99% false alarms. So they just really don't help people protect their families. Maybe it makes you feel better for a little while, but it doesn't actually solve the problem. So I started looking at like what really wealthy people do, and they have people that sit in a guard shack and they watch cameras all day and they pay tens of thousands of dollars a month for that service because it works.
The problem is obviously if you've got a bill for $25,000 a month, I would say there's a very small population that's going to write a check for that every single month. So what we realized was that if we use AI to both make the guards more effective, more efficient, and able to do their job, and then we use all the technologies that are available now to drive that into unit economics, where instead of having one guard or 10 guards protecting one house, we have one guard protecting 10 homes, then protecting 100 homes and now protecting 500 homes and that's really what our business is all about.
Lukas:
Got you. So the vision systems are deciding which cameras are interesting to look at and then a human actually verifies that there's something going on. Is that right?
Dave:
Yeah, and it's raw. That's exactly ... We use the vision systems, which is where deep neural networks have made a lot of progress to choose what to look at. Then we're getting increasingly sophisticated with that to not only choose which cameras to look at, but where to look in the cameras. What are the areas of interest? Hey, if something happened five seconds ago, here are the references that kind of wrap around this event so that you as a human being have the full context, even in the course of like half a second.
Lukas:
So how much of that is interface for the operator and how much of it is like intelligence? Where do you sort of invest your effort?
Dave:
That's a great question. So, because nothing exists like this, we had to build the company in a vertically integrated fashion. So we actually build our own hardware. Behind me we have our hub, which we actually built ourselves and that runs the AI in the customer's home. So we've had to invest a lot of effort. Vertically integrated business is like by far the most complex way to do it. So we have a lot of investments that's gone into AI, we've got a lot of investment that's gone into operations, but really the thesis of the business is, hey, AI can change all of those things if you use AI correctly. This is an AI-driven business because of the nature of what we do, which is video oriented. We can make our operations team smarter. We can make our customer care team smarter. We can make our engineering team smarter by integrating that throughout.
Lukas:
Interesting. So what's the part that runs in the customer's home? You actually do some processing before it even goes to you?
Dave:
Yeah, I'll tilt this to the side here. So right there is our hub, it's right here in my office. Down here are the next generation versions of the hub, which we're working on right now. This is both my home office and our R&D lab seeing as COVID has helped us really concentrate our real estate effort. So what we run in the home actually is almost everything. So we've really focused on moving everything to the edge and there's a lot of work that's being done right now on edge processing. From what we're seeing and what we're building on that curve is just astronomically improving year over year.
What we currently do is we run a version of Linux in the home that has basically all of our BI stack. It's all encrypted and whatnot, but we run almost all of the business logic in the home so that the decisions can be made in real time with the camera, and as you might guess in a security context, the difference between real time and 500 millisecond latency and communication is everything. So we're able to do stuff at just very, very rapid speeds by doing that.
Lukas:
So what exactly is the business decision? Is there like a possible issue? Is that right?
Dave:
Yeah. So if you think about, if you've got a Ring at your house or a Nest camera and you actually turn on alerts, which I highly recommend you don't do, you're going to get like 1,000 different alerts. So the first decision that the AI has to make is, is this alert worth reviewing in further detail and to what detail? Then the second thing it needs to determine is what's in the field of view and is that worthy of taking the time of a human guard to review this? Then the third decision is when the guard is reviewing this, what are all the other pieces of information I need to share with that guard to make sure they can make an informed decision in the course of a couple of seconds?
So again, all of that runs in the home on that hub, and then by doing it in that local situation, all that back and forth communication happens in sub 10 millisecond timeframes versus 20 to 500 milliseconds back and forth to the cloud.
Lukas:
Well, that's really interesting. Can you talk about your hardware?
Dave:
Sure. So the current version of hardware that we're running is a Qualcomm Snapdragon 820. So it's what you'd find in a Samsung Galaxy S6. So if you open up our hub and you peel back all the boards, you're going to see something that is exactly shaped like this at the bottom of the hub and there's a bunch of circuitry around it, but it's literally kind of the reference design for the Samsung Galaxy S6 sitting in the middle and then we put all of our electronics around the outside of that. That was quite an adventure for me because going into this, my experience with hardware had been the robotics that I built in a lab at Stanford when I was there. That was build it once, use it once and make it do everything that it can.
Then a bunch of Raspberry Pis and the types of stuff that you build in your house. That's a robot that I built of a BBH using a bunch of Raspberry Pis and Arduinos and stuff like that. So you think you know about hardware. Just very, very briefly, I will summarize all of my learnings in the last four years. You don't know diddly squat. If that's what you know, you don't know anything about hardware. So we really had to take the time to learn about how do we design that and manufacture that with high quality and solve a bunch of the problems. That last mile of real human beings is really important.
Lukas:
I feel like all of my friends that do hardware, I'm a little jealous, because it seems so cool and they absolutely hate it and they like to complain about it.
Dave:
So I don't hate it. I would say though that the amount of learning that I experienced versus the amount that I expected is a ratio of about 10 to one. Whereas, for most smart people, you go into something, you kind of get a sense of it and you're like, maybe I'm off by 40% or 50%, or maybe it's double. Hardware was definitely a 10 to one ratio of how much I had to learn in order to get into market and be productive.
Lukas:
Interesting. I wanted to ask you actually-
Dave:
It's cool stuff though. It's amazing what's happening in hardware right now though, I will say. Sorry to interrupt you.
Lukas:
What's going on?
Dave:
The Google Coral Board is really phenomenal. The work that Nvidia has been doing is really great. I think my favorite thing though, is just that it's not all about Nvidia. So on the training side, it's still really like an Nvidia dominated world and you can get into China where in China, I see a lot of new R&D happening outside of the Nvidia world, but a lot of that's not really readily available to us here. On the mobile and edge side of the world, you've got everything from, like Rockchip has what they're calling an NPU, that is a neural processing unit and it's an accelerator.
You've got the Google Coral Board, which is driving quantization, which I think is super sweet. That makes things much simpler from a mathematical perspective, but way, way, way, way faster and way lower power. You've got Nvidia doing their Jetson series, but I think overall, what I would say is that in the training side of the world, Nvidia is here and everybody else's here. On the runtime side of the world, Nvidia is here and there are people that I think that are better, significantly better on a cost for performance, and overall performance basis. So we're seeing a lot of innovation happening there and XP, which is another chip manufacturer. They're launching their own NPU in Q1 of 2021. There's just a lot of promise to see a lot of competition, which I think is going to drive a lot of improvement.
Lukas:
Actually, how do you evaluate, why did you decide to go with the Qualcomm board over something else? How do you think about it?
Dave:
So we chose the Qualcomm board, great question, in 2018, and there were three primary factors. The first one was, does it do what we need it to do, which by the way at the time there were a lot of things that would say, like we do AI and they would come with like a pre-trained inception model. If you changed anything on it, it would break. The early days of runtime AI were pretty limited in terms of the scope of what things could do. So that was number one. Number two, was it's a consumer price point. So we're selling direct to consumer.
So while my hub is, that's a $250 cogs piece of equipment. That cost me 250 bucks. If you buy an Arlo or a Nest or any of that stuff, you know this. The end user price is lower than that. So I had to really make sure that it fit within the constraints of something that consumers could afford, and then the third thing was it needed to be size-wise and power-wise kind of containable. So we couldn't just ship people an i7 with an Nvidia GTX card in it. That wasn't going to work either.
Lukas:
Cool. One question I really am dying to ask you. So, I thought of you as a very successful entrepreneur. I actually have this memory of watching you pitch the CTO of eBay many years ago and just thinking like, wow, your presentation is so much better than my presentation. Just wondering, where did you hire this marcomm team that's making this stuff? So that's kind of how I always thought about you. So I was really surprised when I was running into some OpenCV bugs and then I think I found you at forked.
Dave:
Did you really find that? That's so funny.
Lukas:
Yeah. I was like, did you like forked OpenCV or something? I remember just thinking like, is this the same guy that I know? That's not possible. What is he doing? I just was wondering, did you stay technical the whole time or was there like a moment where you got back into this stuff or what's the story there?
Dave:
Well, first of all, thank you for the shameful comment. I'm like literally red in the face here. I appreciate the compliment. I love being technical. I have always had this office in my home, which now is where I live, obviously 100% of my time, but I've always kept this and my wife has always been super supportive of me and it looks like you have something kind of similar in your garage in all the videos I've seen from you of just stay tinkering, stay busy, keep my hands dirty a little bit. There was a period of about three, four years at my last company, which was an enterprise sales company where I was traveling so much. I couldn't be super technical, but I took about a year and a half gap and started getting really, really technical again.
What I found, what you found actually is that there was a version of CUDA, which is the Nvidia library that wasn't compatible with OpenCV. I ended up forking it. I ended up fixing the bugs so that you could run AI and OpenCV at the same time because if you were trying to run AI with OpenCV, you couldn't do it for like a year. So I forked it, I patched it and I figured out how to make it so that you could run AI with OpenCV on Linux, on the most recent version of CUDA and I was blown away.
I did it myself and I was getting like hundreds and hundreds and hundreds of downloads every single day. I was getting bugs submitted to me on the main OpenCV branch and I was like, all right, well, I guess that's cool. It's funny that you ran into it though and for like a year and a half, I was the primary maintainer of OpenCV for all AI researchers around the US. It was pretty sweet.
Lukas:
That's so awesome. I remember looking at your patch and just thinking, man, this is like really deep.
Dave:
There was a bug in the compiled C code that made it incompatible in terms of the data structure and some of the libraries with the most recent version of CUDA.
Lukas:
At that point, were you already working on this new company? Was that what was happening?
Dave:
Well, I wasn't really working on the company. I was working on AI. In fact, I was really fortunate that I got to work with the guys over at Lux Capital. I went to them and just said, "Hey, I'm going to take a couple of years off. I want somebody to bounce ideas off of, as I explore deep learning, because I think it's real." Within a month I had built ... I was using genetic learning, but using a deep learning algorithm and some of the new stuff coming out in terms of vision. So I used a vision feature set using some of the ImageNet competition vision libraries, and then running just a stupid genetic algorithm to play Mario World.
I built the world's best Mario World playing system in like three weeks and I was like, whoa, here I am in my little garage and I can beat the entire game of Super Mario Brothers. This has to be something. I haven't coded in four years and here I am with the state-of-the-art. So I went and I worked with the guys over at Lux and I said, "Hey, can I spend a year just coming in and out of your office and bounce ideas off of you as to what we can do with this technology," because I want to make sure that we're thinking about the specific application in a way that can build a business. That for me, what was most important is that it really had an impact on the world.
Make a dent in the world around us, because as you said, I'm fortunate enough that I can kind of choose the business problem I want to solve and while making great returns for investors is absolutely a necessary requirement. I had the ability to take an extra 18 months and make sure that I also did something that when I look back and I talked to my kids about it, that I'm really proud of and that they are proud of and they can be a part of.
So we ended up spending that time trying to figure out is there really a business model here? A lot of people said security is a bad industry. It's one that's dominated by ADT. It's got all these old players and people buy based on brand and they don't buy based on capability, and I was able to convince them and myself that if you could do a step function in terms of capability, just completely change the game, make something available, going from reactive to proactive, entirely based on the use of this new AI technology that we can change the way people buy.
That's the entire thesis of Deep Sentinel is, screw ADT. It just doesn't matter. Whoever wants to buy that can buy it, but if I can say, "Look, you can go from having the police show up less than 5% of the time, eight hours after your house was burglarized to preventing crimes 95% of the time, isn't that a different reason to buy?" So that was when I was getting technical again, to make sure that I understood what the underlying technology was about, and if I can add one more thing to that answer, my favorite thing I did in that whole period was I started a journal club because I wasn't technical enough to understand the state-of-the-art at that point. I'd been so long since I'd been in AI, like 12 years, 13 years since I was at Stanford.
So I started a journal club and just by offering free pizza, all of a sudden I had these 20 people showing up together and just saying, "Hey, that's a topic I want to talk about too." We ended up building this little kernel of about nine, 10 people that are still going today, five years later. Every single week we meet and we read papers together and we debate things. Because they're all smarter than I am, for literally the cost of one pizza every single week, I was able to get them to read this paper with me and then explain it to me.
I just pretended that I knew what it was about, and then they corrected all of the underlying nature of it. So it really kind of reminded me of the power to just bring people together and learn from each other. Again, I think it's been five years and the same group of people, we still meet every single week. We met last night and we read a paper on Facebook's new object detection algorithm.
Lukas:
That sounds so fun. Where did you find these people that are smarter than you?
Dave:
Dude, I literally just post it on Meetup. I made this thing called East Bay ML. I posted it on Meetup and I started broadcasting it. Actually I did a little bit of Facebook advertising too, like for people that had Python as an interest in their Facebook profile, but I only spent like 300 or $400 and I got hundreds of people to sign up for this meetup. We now we're at the point where every other week we do a paper and then in the alternate weeks we do code and we literally like load up Colab and code together as a group.
These are now, because we're in COVID, about 20% of the people we've never even met them. We don't even know who these people are, but they just show up and we broadcast it. For me, it allows me to stay technical while I'm the CEO of this company without having to kind of distract or bug my engineers.
Lukas:
That's so cool. What kinds of coding do you do?
Dave:
Mostly Python and Colab right now at this point. I would say I do a little tiny bit on CHIP. So I have like a Coral Board and Nvidia board and a bunch of like CPU GPUs in here. So I do a little bit of that, but most of my work now is really in Colab just to make sure I kind of stay fresh on where are we at with neural network architectures and how good do they really perform in reality.
Lukas:
The improvements in ML over the last few years, does it affect your business? Is it important to use the very latest stuff to make your company work?
Dave:
It's not important to make it work. It's important to make it better and in the words of somebody that I think is, again, way smarter than me, Kai-Fu Lee said that AI has moved into the age of implementation. It's no longer in the age of science, it's in the age of implementation. I would disagree with it a little bit. If you kind of look at the S curve of innovation, where you've kind of plateaued out in terms of the actual innovation, I think the neatest thing about AI is that we're at both ends of that curve simultaneously.
We have implementations that are at massive scale, like Facebook doing facial recognition on all of your photos and at the same time in the last year we have seen BERT, which from a language processing perspective is a step function better than all the NLP that came before it. We saw AlphaZero come out, which I think is just a phenomenal leap forward in terms of reinforcement learning.
Then just in the last four months, we saw the new dirt paper come out of Facebook, which is an object detection algorithm that is also, I think it's like 25% better than the predecessor at mean average precision. Then Google had a similar one called EfficientDet, but to see 20% improvements still in the course of a year, that's pretty amazing that we're both in implementation and in the research phases. So for us, what that means is that we have to continually be climbing that ladder because we get operational leverage, we get margin and we get operational effectiveness by continually researching how each of those different new technologies affect our business.
Lukas:
Got it. That makes sense. Has it been an adjustment to go from working on maybe not like safety, critical applications to like an application where you really can't afford to make mistakes or do I even have that right? I would assume that in your world now the cost of an error is much higher than with a recommendation system or stuff you've worked on in the past.
Dave:
It is. I think though, at the same time, I took a card from Elon Musk's deck, which is if you're going to fail, fail early, fail fast, fail hard, and then get to the next plateau. Because my advantage in being a startup again, versus kind of the behemoth of a giant slow moving idiots of the world, not that ADT are idiots, but let's just call them idiots for the purpose of this conversation. That you only have really one advantage, and that is your willingness to just look the monster in the eye and fail fast and fail early.
So yes, it is a mission critical product and we absolutely strive to be better than everyone else, but we also recognize that if we're going to make these mistakes and do this learning, we've got to do it in this phase of the business so that when we get to the hundreds of thousands of customers, we're at another plateau.
Lukas:
I guess if ADT, if it's true, I wasn't aware of this. If ADT actually, when the alarm goes off, nothing typically happens, then maybe it's okay if occasionally your system has alarm go off and nothing happens, is that right? Or that's probably not what you put in your marketing content.
Dave:
I would not say that it's all right if we don't succeed, but I will say that even in the really, really rare cases where we don't operate as quickly as we would like, we're still a hundred times better than the next best alternative. We have customers where we've messed up and we have, and that happens but on average and in fact to the 99th percentile, they understand because they understand that when they went to market and said, "Hey, who are Deep Sentinels' competitors?" The Google result list is zero. There's literally nobody else that can do what we do at a consumer affordable price point, and nobody else that does it as effective and as at much scale as we do. So they are generally pretty understanding that like, I get that you're learning. I want you to protect my family. I rely on you. I can't put up with this, but I understand.
Lukas:
Interesting. It's funny, I've tried to ask this question to a whole bunch of people and I think I've kind of stopped but I always wonder, with people working on autonomous vehicles or applications where there's really like, safety is a huge issue, it just seems like incredibly hard to know. I've looked at so many ROC curves in my life and pick where you go in the precision recall trade-off, but I'm so glad that I'm not in your shoes trying to pick the precision recall trade-off. Do you have anything to say on how you-
Dave:
That for me is a little bit easier. So on autonomous driving, you have to choose left or right. Faster, slow, which is actually a more hard question than ours, which is show this to a guard and allow the guard to review it. Because I have a trade-off. I can have more guards than I need, and I can show them more video and I can even show them trash videos for a period of time, because I can afford to do that. As long as I can maintain the fidelity of their operational behavior. I can actually just solve that problem with money. Whereas if you're choosing left or right, you got to choose one, you can't choose both. That's, I think a much harder decision that you really can't solve with money. So they have to go slower. I can go faster and just lose a little bit more money for a short period of time.
Lukas:
Although you can't do that infinitely. So I feel like that logic would apply anywhere. So at some point you have to tolerate some amount of error. There's some amount of error you always have to tolerate with these systems. So how do you even think about that?
Dave:
There is. So the way that I think about it again, because our decisions, the core decision, the most dangerous decision that we make is binary, what you can do is you just perform it as an experiment. You operate at the lower threshold, but then you measure at a higher threshold and you just measure the gap here.
Lukas:
Got it.
Dave:
Then you have a level of acceptable gap, and then the second thing that we do is we also measure that across the life cycle of an event. So because we're looking at events that are, let's say 25 seconds long, that's 500 frames. So I actually have 500 opportunities to make the decision to go wild with that event. So I don't have to make that decision at this point. Now, I can't tolerate it if it's 20 seconds late, but can I tolerate it if it's 300 milliseconds later? I generally can't. So what we've done is we've been able to drive this threshold where we keep that number pretty much zero, by the way, to be honest with you. Then we drive other dimensions of flexibility in our decision making, instead of driving the decision of I'm going to let this event go, because we don't consider that to be acceptable and we keep that number pretty much zero.
Lukas:
Do you pass along your confidence value to the operator?
Dave:
Great question. No, because in general, what we have seen is that the confidence values, and you see this all the time. I'm 99% sure that my cat has a dog. Like it's not real information. I don't find, and the shape of those softmax curves are so sensitive and the number that comes out is not a true confidence number in the sense that humans interpret them. Like 0.6, which would turn into 60%, really doesn't mean 0.6.
In fact, if you look at the shape of most of the softmax curves, they're very heavily weighted on the zero, between zero and 5% and between 90 and 1005. There's almost nothing that's in that useful range in the middle. So we tend to just trim it off because then if you try to normalize it so that it's distributed evenly between zero and 100, what you're doing is you're taking basically 0.9 to 0.91 and making that this huge section between 50% and 75% confident, and we find that it's not actually that representative, like the granularity doesn't really exist in the confidence, at least in the curves that we've seen.
Lukas:
Interesting. Do you ever go back and deploy new models to your customers, like existing customer?
Dave:
All the time. We deploy new models on a weekly, monthly basis.
Lukas:
Interesting. Will you train them on the customer's data?
Dave:
Yeah. So we train them both on the customer's data in aggregate, and then we actually have a patent on personalized training at the edge or a patent pending that we use the individual customer's data to fine tune the model in their home, and that's one of the neatest things about having the hardware in the home is we have a huge chunk of data in the home and we also have a model in the home. So there's a bunch of advancements that have been made in semi-supervised learning where you can refine these models at the edge. So we could do that both in terms of how we interpret the final coefficients coming out that fully connected later, as well as do that on actually retraining some of the weights in the model in the home.
Lukas:
So it'll actually retrain on the edge and not phone home?
Dave:
Yeah so we can do some amount of that. We have tended to do that only on the fully connected layer or like maybe one or two layers and we've done a bunch of tests to see, like where does that matter. In general, we found that even just the fully connected layer, rebalancing the way to the fully connected layer is pretty effective at driving a massive shift in mean average precision, especially when it comes down to our problem, very specifically is we have stationary cameras. So a big part of stationary cameras in terms of a vision problem is identifying the background. So when you have hundreds and thousands and millions of images of what a camera looks like on average, you can develop a very clean sense of what the features are of the background, so that your background subtraction becomes even more effective across the board.
Lukas:
Do you have any examples of customers that have kind of taken a camera and deployed it in some way that you didn't expect that caused the system to struggle?
Dave:
We have tons of examples where the interface with the wet part of the world, the human being part of the world is wackadoo. We have to have a very specific policy about what you do when you see naked people, for example. So we have customers that we have discovered that ... And you know him as well. So Sean Parker used to always say, "The world is broken up into two categories of people, exhibitionists and voyeurs." I'll tell you a product that has people watching your cameras, you definitely find people in that first category. So we have customers that will put their camera in their house and then proceed to treat it as if it were an adult rated YouTube channel.
Lukas:
I'd have to ask, and maybe we should take this out, but ... What do you do?
Dave:
We turn the camera off and we apologize to the guard that had to watch it. So we have special features where that escalates immediately. We disable it. We send them a very nice letter that says, "We noticed that you've installed your camera indoors. We have gone ahead and disabled that. Please verify when you've moved that outdoors and let us know," but it's a product that touches the real world. This isn't a Nest or an Arlo where it's a dumb device just recording. This is a device that is a live ecosystem. So we find all kinds of stuff, and this is one of the things I really love about the business that we've started is that we have this huge, deep technical investment in AI.
Then we have all of these different real scenarios that it's being trained against that when you compare our AI against what you would get from a horizontal vendor that does image recognition and classification, there's no way that they've spent the time to say like, "Hey, what does it look like when you're inside and there's a guy dancing in his underwear in front of your camera?" Nest doesn't have that, really. They don't do that. So we've just solved a lot of these really neat technical problems that are at this interface between cameras and human beings that I think is really interesting.
Lukas:
I guess this is outside of AI maybe, but I'm kind of curious as another entrepreneur to entrepreneur, it seems like you have this interesting challenge of proving to your consumers that your camera's actually better. I feel like a lot of things it's like, obviously better, but I guess in security, it's better is the absence of issues. how do you demonstrate that your camera's better?
Dave:
It's more of a marketing question in some senses. So one of the challenges that we have is that for the last 30 years, the definition of better for cameras is pixels per inch, or color resolution where what we did is we changed the game. The camera really is a dumb camera. Like our camera is great and it's as good as anybody else's, but it's not better as a camera. In fact, there are areas where the new version of Nest, new version of Ring or whatever are technically better than our camera, but none of them do what our camera does. It's more about the capability. So what we found is that we break that problem down into two challenges. We have to handle the matrix of like, are you better and are you worse?
Then the second piece of it's really what you said, which is how do you deliver peace of mind, which is such a dangerous word from a marketing perspective. So what we've really focused on, we have the series of videos called the stopped videos, that show us just stopping crimes. Boom, repeatedly over and over. They're kind of individually boring because the point where the guy walks up to the front of a 7-Eleven, we have this 7-Eleven that gets burglarized literally once a week. Guy walks up with a crowbar and he swings at the door and that's the end of our video, because our guards get on and say, "Hey, jerk, get out of here. The police are on their way," and the guy walks away. Whereas if you had a dumb camera, you get this really cool video that for the next 45 seconds, you see this guy banging on a window.
So what we've had to do is we've had to train the market that like, Hey, prevention is possible. It's much harder to sell proactive in some senses. So we've really spiced it up and made it exciting and made these video series about it. That leads to the question, "Hey, how do you guys do that?" We now have an AI, an autonomous AI called AI Deterrent that triggers within 200 milliseconds of the AI detecting somebody suspicious at night.
That's pretty sweet too, because we're literally intervening in these crimes even before it would get to a guard and that typically buys the guard another three to five seconds where the person's talking to the computer and then the guards like, "No, man, I really need you to go."
Lukas:
Do you worry about adversarial attacks? Like someone figuring out your algorithm and then finding ways to go in undetected? I would think you might be too smart for that to really be an issue, but maybe that is.
Dave:
I think, yes, we think about it. I think about that all the time, because again, my business is security. So that's the way that you have to view the world when you're insecurity but from a business perspective, at the point that we have people really developing adversarial attacks, we're at another stage in our development.
Lukas:
Got it. Well, another totally different thread that I wanted to ask you about, because people ask me about this all the time. You're a fairly active angel investor and long time, super successful entrepreneur. I was kind of wondering how you think about investing in AI companies. Like, what you look for there.
Dave:
So I have actually made the decision about four or five years ago to stop angel investing for the most part because I did okay on some of them, but I found that the people that are really good at angel investing do that full time and that I didn't want to do it full time. I find it really neat to meet with entrepreneurs and hear their story and like you. I like helping, I like being there for them and the investment piece was, in some senses, clouding that interaction just because I don't have enough money to invest in every one of the cool entrepreneurs that I see, and at the same time, I don't have enough time to spend time with a bunch of them.
So what I ended up doing was I ended up joining a venture firm as a venture partner so that they could deal with the investment side of things and I could just really focus on the advising and talking to them. What I've generally seen are kind of two camps of AI companies, and I think this is what you always see with emerging technologies is you see the geeky technology oriented founder who really doesn't understand the business that they're in and they're just like super smart and they're endearing.
You want to help them, and then you see the smarmy, sorry for all the guys out there at girls out there that are this, but like the sales person. That shows up and they're like, "And the AI is just amazing. You would not believe it's using this thing called stochastic gradient dissent." You're just like, "Oh my God." So after I pulled a knife out of my eye, I don't advise those companies, but I entirely focus on those founders that have that chutzpah.
They've taken the risk to do something that they are not good at. I am going to go and start a business as a tech person, and I've got this crazy great insight and I find that to be much more compelling and interesting and fun for me versus trying to educate some sales schmoe on what's really happening under the covers.
Lukas:
I guess you've watched a lot of technical founders become wildly successful. You've been doing this a long time and have had really front row seats to Silicon Valley. We should put your bio in the show notes, but it's impressive. Are there any patterns that you've noticed over the years of like, who you meet in that mode as like a technical founder? Because I'm sure that's a lot of the people watching this and listening to this and how they actually succeed, like which people are likely to succeed and then what they do to make themselves successful?
Dave:
So the two things that I think are pretty consistent is embrace crazy, consistently. The things that other people aren't doing, that's exactly the world that you have to live in. You can't do all of crazy, but if you live in just not crazy, then you're competing with them on their terms and that is consistently a path to failure. So you have to embrace crazy. Then number two is the technical founders that I have seen get wildly successful, they aren't necessarily really self-aware, but they're sufficiently self-aware to hire their compliment. That's either in like one really amazing person or in two or three or a whole team, but they find some way to remain the crazy person and have a team that embraces that and supports them and gets the other stuff done.
So I want to talk about social responsibility, because I've seen just this market shift in the role of data in our society over the last six months and not necessarily in a positive way or a negative way, just a standout observation that if you look at the two or three big things that have happened to the United States and the world and in the last six months, the COVID obviously, and Black Lives Matter, both of those have been really hard for our society to get a hold of because we have built a society that is headline based, whether you blame it on Twitter or whatever, I don't care.
I'm not going to blame anybody. It's just that we are a headline-based society and the problems of COVID are so incredibly complex. They are deep statistics, they are statistics that don't lend themselves and they're so data oriented. They don't lend themselves to a quick 40 word summary as to what's going on. The same thing with Black Lives Matter. You could find a black person that makes more than a white person doing the same job. Absolutely, and if I wanted to, I could write that article up, ship it off probably to Fox News and they would run it.
I could also find a black man that makes $100 per hour, less than a white person at the same job, ship that off to MSNBC and they would run with it. That instantiation based, existence-based proofs are exactly the opposite of what statistical distributions are all about. A statistical distribution is about capturing a totality of a problem and COVID, I think brought the concept of an exponential curve to the masses in a way that, I think 16% of Americans get through calculus and statistics in high school, but now 100% of Americans have been exposed to the concept of flattening the curve and why an exponential curve is important.
Then we got exposed to, unintentionally, what happens when you intervene in an exponential curve? Oh well, but that never was going to happen. Well, now we have intervention-based statistics that now, I'd say less than 16% of Americans are aware of, but statistics have moved from being a fringe thing that affected insurance and financial markets to being something that impacts all of our lives. I think that's something that is both exciting to me and scary because we, as data scientists, have generally lived in this world of like, I'm just living in the numbers. So whatever I say is just the numbers and I have no social responsibility, and I think that's absolutely wrong, that that is categorically an incorrect statement. In fact, I think the opposite is true. It's that numbers have become so important in our society that it's our job to go to primary sources.
It's our job to educate our friends. It's our job to read past the headlines and play an active role of helping people understand what a distribution means to those things. I had a couple of examples, for me, that really stood out. At the beginning of COVID, there's a great blog called Towards Data Science. I don't know if you read that one, but I like it. I think it's pretty good.
Lukas:
Totally. Terrific.
Dave:
There was an article written in Towards Data Science that said the cancellation of these conferences is absolutely the wrong thing to do. In fact, I'll just share it for those of you that are watching. It's this article here. He invented this term called Coronoia or like paranoia. Coronoia and no slight to the author because I think we were all learning.
So I don't want to slight this author, but I hated the article. The reason I hated the article was because the paper that the article states, here's the population of Spain. It's 46.66 million people. The total cases of coronavirus in Spain is two. So therefore the number of people that would get coronavirus at this conference is 0.0046. What I really hated about this was that we had an opportunity to have an educational moment and have a conversation, and instead, what we did is we chose a false population. The population of Spain and the number of cases in Spain is not the problem. We simplified the problem in a way that abstracted out the exponential growth component and the exposure component and we stopped and didn't believe in intervention.
We didn't analyze the actual data, and this is the danger of statistics to me is that you can use statistics to justify something. If you just trick people by choosing the wrong population, you trick people by choosing the wrong coefficients, you trick people by using the wrong underlying model. In this case, he used a linear model instead of a exponential exposure model. That to me, I think is the quintessence of what we as data scientists must tackle is to be ... Sure, we can be opinionated, but to be balanced and I think it's just so much more important than it ever has been because we, as a population, don't have the educational system, so that 100% of Americans understand what statistics mean.
Lukas:
That reminds me that there's been, I think another observation that a lot of folks have made about machine learning algorithms that when they're ... Based on training data, that could have underlying bias in that data. We can easily end up with models that reinforce society's biases and make it worse and actually, you make a device that calls the police. So how do you think about that in your device, in your training data and what you do?
Dave:
Dude, such an important question and I just made this big statement that we have to be proactive. We have to recognize that if we just say, "I'm agnostic because I live in the data. So therefore I'm not biased," is a falsehood. Mark Zuckerberg said that in front of Congress and I loved that he did that because it shone a bright light on the fact that that is a lie. I don't know that Mark did that intentionally because I think we all believed that right up until very recently. I don't think we were exposed to our own biases and our own potential for bias until that happened.
What's interesting about that moment is I happened to be in Washington DC on that exact same day and I had called a meeting with the ACLU and the NAACP and a number of other civil rights organizations. I presented my business to them and they all said the same thing. I said, "No, no, no, no, no, no, we can't be biased because we're based on the data." What it did is it really forced me to take a step back and say, "Actually, could I?" Let me ask the question.
Instead of making the statement that because I'm focused on the data, I can't be biased, let me make the hypothesis. Let me make the null hypothesis that states, because I'm based in data, I can't be biased and then disprove it. I found that there were just hundreds and hundreds of ways to disprove it. So we actually designed our system specifically to be race blind. We intentionally designed it to be raised blind. We manage the training data coming in so that we can identify people.
Then the second thing we did was we used the data in a way that cannot be abused from a race perspective and an ethnicity perspective. Then the third thing we did, which I did not expect to do. In fact, I came into that meeting with the NAACP saying, "We're not going to do this." I said, "We're going to not track the race of the people that we call the police on." They said, "In fact, I want you to. I want you to do that, but I want you to store it in a system that's not being used for machine learning. I want you to store it in a system that you use for auditing your employees and you use it for auditing your own business to make sure that you are doing the right things." So we did those three things and I feel wildly good about what we did because one, we were proactive.
We did it before we got in trouble. We did it before we called the police on the wrong person. Two, it was an educational experience for me. Again, I was very strongly in the camp. I was in a recommendation system business. I built the recommender at Amazon. Those systems are entirely based on, hey, you just use the data and whatever the data say, that's what you optimize to. I was surprised and pleasantly pleased with the results of taking that step back and saying, "Let's treat that instead of as a conclusion, as a null hypothesis," and I learned a lot.
Lukas:
It's a strong claim though, that your system can't use race. How do you know that's the case?
Dave:
So it's not that it can't use race. It's that we make sure that the distribution coming in on the training side is designed to be race agnostic. It's a fair distribution on the input side, which requires tweaking the distribution on the open side. Then the second thing is we don't allow the system to perform specific activities. We don't do facial recognition at this point because we recognize there are some issues with facial recognition and we don't allow it to call the police. We focus on classifying something that is itself, race agnostic. So we focus on behavior and we focus on classifying people versus cars. We do not focus on classifying suspicious looking person, and we've specifically designed the system to not be able to do that, if that makes sense.
Lukas:
So the system looks for people, not suspicious people.
Dave:
That's right.
Lukas:
Then if it sees a person-
Dave:
Then it identifies the behavior and it separates the behavior identification from the person identification. This is what I mean by actually designing it to not be able to do that. We do not allow the behavior identification portion, the suspicious identification portion to know what the person looks like. It does not have access to the pixels. It only has access to a completely removed representation of the behavior. So it's a set of vectors and features that have been completely ... Have entirely removed all the pixels.
Lukas:
So it synthesizes the pixels into some information that then-
Dave:
That's right. So we created an intermediary data structure, which is here is this object, here is its classification and here are the dimensions of its motion and then we drew a hard line. That system does not talk to the other system that says, based on these motion vectors and this description of behavior, this is suspicious or not suspicious.
Lukas:
Interesting.
Dave:
Again, it was kind of non-intuitive to do that because as a data scientist, I would say, I just want to go from here to here. That's going to be much more accurate, and from a mean average precision perspective, it might be, but it also is exposed in a real world context to having bias in the end to end but by enforcing this intermediary that gets rid of the pixels that might contain anything having to do with race, it makes sure that it can't propagate through.
Lukas:
Then you collect data that tells you that it's not using race, you're saying, or how do-
Dave:
So what we do is we then collect data in an independent system that says, these guards were calling the police on only Latina people, or only on Asian people, or only on black people and making sure that we are auditing all chunks of the system against the racial distribution, so that they're acting in a way that is consistent.
Lukas:
Is there any other groups that you worry about? Do you look at gender too or other aspects of appearance or anything like that?
Dave:
We do. So we track the race, the gender, and then generally the age as well, because age is a really important one. So one of the things that I learned in my meeting with the NAACP is that black males under 18 are frequently classified both by police and by witnesses as being adults and you need to treat minors differently. We have a social responsibility to say, "Hey, man, I see you TP-ing that house. Get out of here."
We call the homeowners and say, "Hey, there's somebody TP-ing your house," instead of, "Hey asshole, I'm calling the police. You need to stop right now." That creates, as we're seeing, the intervention of escalation creates a different outcome. The intervention creates the outcome. So you better darn well, choose that intervention based on real data and if at the end of the day, black males are specifically classified by people frequently as being older, then we got to train for that, we got to compensate for that. We got to treat minors as minors because their brain development is the same regardless of race. We have to enable them to develop their brains.
Lukas:
Have you written about this at all? Is there some place we could point people that wanted to learn more about this? I'm just not sure we can cover all the questions that-
Dave:
I haven't. I haven't written about it as much as I probably should have. I actually just did that. I had that meeting with those groups and I just did that for the purposes of ourselves. I probably should, at some point, write about why I did that and what it means and what the social impact is of that. I did do a little letter to our customer base, when the Black Lives Matter movement was really taking hold and I said, "Look, I want you to know what we've done, because I'm really proud of it and I don't want racism to exist in our business because it's so dangerous." Because we live at this intersection between human beings and the police. We have a real responsibility to take that at a high level.
I'd say 99% of our customers were like, "That's awesome that you've done that." You don't have to pick a political side to say, "You've done the socially and business responsible thing." Interestingly enough, we did get two or three customers that came back and were like, "F you, you're making this up."
Lukas:
Wow.
Dave:
Which was surprising to me.
Lukas:
So the hostility was like, you're making this up or we don't like that you're doing this?
Dave:
You're making up that you did this, you're making up that there's a real problem. There is no problem. Ben Shapiro told me that there's no problem.
Lukas:
Right, right. Got it.
Dave:
At the end of the day, I very much strongly believe in the First Amendment and I love that people have their opinions. Again, this is where statistics come into play. There are no real statistics that say that our country's not racially biased. Like zero.
Lukas:
Well, let me ask you one final question, maybe a little less intense than that one. In your process of going from a model to a deployed system, were then the surprising bottlenecks? I think everybody senses that it's hard, but what were the points where like this took a lot longer, it was a lot harder than you were imagining that it would be?
Dave:
So number one, first and foremost is specific operands on specific chipsets. So for example, Qualcomm implements this huge stack of operands and it implements them in a way that you can accelerate. So as long as you stay in that little pool, you're okay, but all the new architectures typically are using new operands. That's a big piece of the new papers is they're getting to state-of-the-art by using this new version of ReLU or this new version of a generalization operand between layers and more activation functions, other activation functions that are non-ReLU based. What we have found is that in the last three years in general, that that problem is being solved with a big baseball bat. Instead of being solved with a surgical tool and being precise, it's being solved with broad swing.
So for example, Qualcomm has the problem that if you don't support it, it just crashes. That's sweet, good to know. Then you've got TFLite, which I think is a much more robust architecture and I like TFLite a lot, but what does TFLite do? Well, if you're using an operand that isn't supported on your particular architecture, we just move everything from that point forward into your CPU. So you go from a model, you make this tiny tweak to a model and it runs in 25 milliseconds, and then you make a tiny tweak and now it takes 750 milliseconds. It's not this kind of like nice smooth curve where things get reconnected in the middle. It's just, and it blows up and as you might guess, systems that are based on an assumption of performance between 25 and 50 milliseconds do not perform well when performance goes to 750 milliseconds.
So I think we've improved in that like TFLite doesn't crash and totally die, but actually at the end of the day, your system crashes and totally dies. So you get the kind of the same outcome without having a system error, which is maybe better and moving in the right direction but certainly not there. It's definitely still a lot harder than I would like. I would say that the second thing that I've learned is how incredibly distinct the world of training is from the world of runtime operations. If you're running it in the cloud and you're willing to pay the GPU prices on the cloud providers, then it's not that big a deal.
Any of the more precise architectures, the level of OS tuning that you have to do, the level of firmware driver management that you have to do, it's a lot more than I thought it was. Again, coming from the Raspberry PI tinkering type part of the world to actually implementing it and having it run 1,000 times an hour, every single hour, 24 hours a day for 365 days a year in all 50 states, that last mile was much, much more complex than I expected.
Lukas:
Interesting. Cool. Well, thanks so much for your time. I really appreciate it. It was great to talk to you.
Dave:
It was great to catch up with you. I wish I could see you face to face.