Transcript 0:00 Hello everyone, and welcome to episode eighty-eight of the IoT For All Podcast. 0:03 I'm your host, Ryan Chacon, and today's guest is Perceive founder and CEO, Steve Teig, who joins us to talk about how Perceive is bringing the next wave of intelligence to IoT through machine learning at the edge. 0:14 To give a little background on Steve, he's an award-winning technologist, entrepreneur, and innovator, has over three hundred and eighty-eight US patents. 0:20 He has been the CTO of three EDA software companies, two biotech companies, and a semiconductor company. So overall, this is a fantastic episode with Steve. 0:28 I really enjoyed our conversation, so if you wanna learn more about the intelligence side of IoT, how machine learning is playing a role, learn more about what Perceive is doing and their chip, Ergo, please do not miss this episode. 0:40 But without further ado, please enjoy this episode with Perceive founder and CEO, Steve Teig. Welcome, Steve, to the IoT For All show. Thanks for being here today. Happy to be here. Yes, it's great to have you on. 0:51 Um, you wanna start off by just kinda giving us a quick introduction, background, experience, anything you think would be relevant about you a-and your, and your history of, um, kinda getting where you are today, just so our audience has a better idea and understanding of who they're listening to? 1:05 Sure. I, I am a serial entrepreneur. 1:09 I've been starting companies for about thirty-five years, have started six, uh, uh, two of which were in software, two of which were in biotechnology, two of which are in hardware, uh, computer hardware. Wow. 1:23 Uh, and on the way, I was the chief technology officer of Cadence as well. Very, very nice. That's a, that's a wide range of companies and industries. Um, sounds very fascinating. Can you talk about kind of f-f-f... 1:36 the, your, the s- s- history or I guess a story behind, um, founding Perceive and, and kind of the problem you all saw in the market? 1:45 I know it spun out of, of the parent company, but just talk about kinda how that process worked, where the ideas came from, how it eventually turned into its own business, 'cause I think that's a very interesting topic. 1:56 Uh, sure. So I was the chief technology officer at Xperi starting about five and a half years ago, and Xperi has a wide array of products in imaging and in audio and so on. 2:08 And w-what I saw was, uh, the opportunity to make the edge in general smart. 2:15 What I saw was that while you have all of this machine learning that was starting to emerge five and a half years ago, uh, i-in people's clouds, in the giant data centers of the world, that the, a, a, a tremendous opportunity that wasn't being exploited is what if you could make gadgets themselves be intelligent? 2:36 What if you could make the- Right... gadgets in your house, uh, understand their environments? And that's what motivated me to pursue that technically and eventually to spin out Perceive to pursue it for commercially. 2:47 Fantastic. 2:48 Can you talk about what Perceive actually does, like, kind of more at a granular level, and how you engage with customer, maybe also what are your target audiences, you know, kind of w-what business and role you're basically playing as it connects to the IoT space? 3:02 Sure. 3:02 So, uh, Perceive builds both, uh, chips and m-machine learning software for edge inference, basically to, uh, provide data center class accuracy at the l-low power that edge gadgets like IoT gadgets, uh, require. 3:23 So the kinds of applications that we're going after are everything from, uh, doorbell cameras and home security cameras to, uh, home appliances to toys to phones. 3:36 Uh, wherever you would reasonably have a sensor, it would be cool to make that sensor understand its environment without sending data to the cloud, and that's the problem we solve. 3:46 And is, are you mainly focused on the consumer side, not, you know, venturing too much into enterprise, or is there a mix there? Uh, it's a mix. So we, our initial foray has been in consumer. 3:57 Uh, we, our first customers are in home security cameras and doorbell cameras and so on, but we're definitely interested in pursuing both consumer and enterprise. So, so then what, 4:09 what makes kind of your offering unique to the market and compared to, I guess, devices or, or, um, a-you know, assets that are, maybe do not have your technology inside them? 4:23 What would the experience look like or how would it differ from a benefit standpoint for someone out there who's listening to this and trying to understand the value provided by having a, a device with, with your technology? 4:35 Well, basically today, if you wanna have a gadget that manifests some intelligence, you have two bad choices. Okay. [laughs] You can either send all the data to some- to, to somebody else's cloud- Uh-huh... 4:45 and give up your privacy so that somebody else's cloud can deal with it, and it's expensive and costs lots of energy. 4:51 Or you can have some toy little chip in your edge device that's low power enough it can run in an edge device, but then it doesn't have the horsepower to give you answers you can trust. 5:01 What we do is fix that problem by providing data center class accuracy, the same kind of sophistication that you expect from the big cloud providers, but at low enough power that we can run at the edge. We are 5:17 20 to 100 times more power efficient than anything else on the market right now, and as a result, we can do world-class inference, but in the context of a battery-powered device running at the edge. That's incredible. 5:30 That's fascinating. 5:31 So, um, are there cu- like, public customers that you're able to reference that you work with, or maybe any use cases that you could take our audience through, maybe in different, different kind of, um, industries or areas that they would be able to understand and kinda see this, I guess, visually in their mind of, you know, how this works and, and what you guys are providing to the market? 5:51 Uh, sure. So we haven't yet announced, uh, customers publicly. Uh-huh. We have several, uh, and, and you'll see their products get announced relatively soon, 'cause we're already making production silicon and so on. 6:02 But, but certainly I can speak to the applications, and I'll speak both present and future. So present are, again, uh, doorbell cameras, uh, home security cameras, uh, uh, appliances. 6:15 But as we look forward-Being able to put really serious contextual awareness into edge gadgets opens up all kinds of applications. 6:25 For example, you know, when I walk up to my microwave oven with broccoli in my hand, why shouldn't the microwave recognize me and say, "Oh, look, it's Steve, and he has broccoli in his hand, and I know how he likes his broccoli cooked. 6:37 I'll just set myself accordingly"? As opposed to having a whole bunch of buttons on the microwave that nobody knows how to use, which is where things are today. Right. 6:45 W- w- why can't I, uh, why can't I have my Roomba, my home vacuum cleaner, uh, be able to tell it, "Would you tell me where I put my loafers? 6:53 I know they're in the house somewhere, but I don't remember where, where I put them," and have it go r- run around the house, figure out where the loafers are, and come back and tell me. 7:01 Why shouldn't we be able to do that? 7:03 Why shouldn't you have your f- your, an, your phone always on, always watching, always listening, but without the creepiness of sending all that data to somebody's cloud, and yet still be able to, to, to, to see and hear and understand so I can interact with it without having to pick it up? 7:20 How about on the, um... You mentioned something earlier with kind of the smart doorbells and things like that, maybe video doorbells, um, and, and potentially video chatting was to attach the phone. 7:31 How, how would those kind of play in by, with your technology? What would they enable a, a consumer to be able to do? 7:37 Well, for example, you wanna be able to tell the difference between the FedEx delivery person on your porch versus your aunt who's visiting, so it should be able to inform you who is on the porch. 7:49 It should be able to tell the difference between delivering a package and picking up a package. 7:54 So it, it should be able to recognize ambient sounds, so if, God forbid, somebody breaks your window, it can tell you, "Somebody just broke your window," and recognize the sounds in its environment. 8:06 So we can put the real intelligence into that edge gadget, where the gadget itself still has, you know, a year of battery life or two years of battery life, uh, while understanding what's going on around it. 8:19 So, so right now, I guess, to kinda educate consumers who may not be as familiar with this technology or these devices, how are, um, kind of, I guess, doorbell, video doorbells now behaving, um, and obviously they're not allowing the kind of, um, intelligence that you all are putting into doorbells. 8:39 So what's the experience like, and kind of what are you, what are you hearing from consumers on, on, on, I guess, where that is maybe not meeting their needs? 8:47 Uh, well, so what's happening today is all of those companies send all of the data directly to the cloud, so you have a fire hose of video and audio that's being sent to somebody else's cloud. Mm-hmm. 8:59 So first of all, there's a violation of privacy and security and all that stuff. It's a little creepy. Um, but the second thing is look how much power you're consuming to send all that data to the cloud. 9:10 Look how much money you're spending to rent storage in the cloud and rent computing in the cloud, and after all that, you still have lots of false positives and false negatives. 9:20 That is, it, it, it hollers when the tree in front of your porch waves in the breeze and says, "Oh, my God, there's a burglar," and yet can fail to recognize [laughs] the burglar. Right. 9:31 Um, it, so, so, uh, w- we feel like we can address that collection of issues by putting really massive horsepower right at the camera so it doesn't miss anything, so, so it, it doesn't have the power cost or the delay of sending something to somebody else's cloud, so it's much less likely to miss what's going on while still having enough computing sophistication to recognize people and animals and vehicles, et cetera. 9:58 Mm-hmm. 9:59 Now, from a sensitivity standpoint and just, like, a user man- or u- user in, um, the way the user would interact with this, is this something that would be able to be controlled by the end user of these devices as far as setting its sensitivity level, like, uh, understanding, um, you know, 10:17 different kinds of thresholds for it to kind of, um, have, have set up for a user? 10:23 Or, or how would, how would that work, or how do you envision that working if, you know, when this comes to, to be used more on a larger scale? Right. So all, all, all of those things are most definitely possible. 10:33 So these are things that will be between our customers and their customers, so the doorbell camera manufacturer and so on. 10:40 With that said, from a, a, an underlying inference point of view, from Perceive and Ergo, our silicon point of view, you absolutely can control, uh, how sensitive it is. 10:51 You can identify regions to say, "I don't care about things in the street. 10:55 I care only about things in my driveway and on the porch, and, uh, y- y- I want it to be more sensitive at night than I am during the day," uh, and so on and so on. 11:06 All of those things are, are easy knobs to tune, uh, even on the end user side. Okay. So let's pull away from these use cases here for a second and talk more about your all's approach, um, to machine learning. 11:17 Uh, at a high level, how would you kind of describe your all's philosophy on machine learning and how it differs from those other ones in the market? 11:23 And, and what does that approach and the stuff you, and, and the solutions and products you're bringing to the market, uh, mean for the scale potential of consumer devices and the IoT industry, uh, just as a whole? 11:38 Uh, fantastic. Well, um, we realized that in trying to go for the 20 to 100x power efficiency advantage over the state of the art, that we really had to start with a clean sheet of paper and frankly to start with math. 11:52 That is, rather than just take the techniques out there and try to fiddle with them, we, we, we really started from first principles to say, "What would a, an inference engine that ran at super low power have to look like? 12:05 How is it even possible in principle to do this?" And we ended up in a pretty new place. We have developed some new mathematical techniques. 12:13 Uh, we, we represent the neural networks, which are the technologies to do the inference, in a different way, in a new way, in a proprietary way that's significantly more efficient than what's out there, and we have mathematical rigor behind all that, so we understand why our methods work and how our methods work and are getting hundreds of X of compression.Versus what this, what, what the, uh, mainstream world is doing out there with inference. 12:42 Um, i- in, in terms of the implications it can have in the marketplace, 12:47 basically the possibility of having the sophistication of data center level computation in something that burns only 20 milliwatts opens the door to having every sensor, wherever your sensor might be, if that's in a phone, in an appliance, in a toy, in your car, wherever, to have every sensor understand its context. 13:08 It can make every, uh, gadget you interact with behave the way it's supposed to because it can understand who's looking at it, who's talking to it, what's going on in the environment, and, and, and have the interfaces reach a level of sophistication that's simply impossible without that context. 13:27 That's fascinating. So, uh, what is the, um, what is the name of the product that you guys have in the market now or you'll be launching soon? 13:35 Uh, I believe Ergo is correct, and kind of if you could talk a little bit more about what Ergo does and, and how it operates and kind of what it enables from a feature standpoint, I think that'd be really interesting to hear about. 13:46 Sure. So indeed, we have, uh, a- announced our chip back in March. We announced our chip, which is called Ergo. 13:53 It's tiny, so our chip is in a seven millimeter by seven millimeter package, so small enough to go in even a wearable. Um, it doesn't require any external memory or external processing. 14:04 You basically can hook sensors up to it, and it can do inference. It can detect faces and people and voices and, and, a- and, and, and all of that stuff. Um, while we have developed 14:17 a great deal of machine learning technology to, to create the models to do object recognition, uh, the embodiment of those models from a customer point of view is in our chips. 14:30 So we, so we're providing chips and applications that run on those chips so that our customers can make their gadgets smarter using Ergo inside them. Very interesting. 14:39 Okay, so take me through a typical customer engagement, um, somebody that would come to you. Is it, is it all... 14:45 Is that how, I guess, to start, is that how it often is, is somebody comes to you through, um, you know, the marketing and the efforts that you would do to educate the market on the, the products you have, and then you work with them to kind of figure out how to implement the technology into their hardware? 15:04 Or is some- is it something that you all are going out and actively finding, um, a device or, or I guess, um, product companies to get your chips into? 15:16 Or what does a typical engagement look like, and at what stage of the development are companies coming to you, and is, is it something that requires them to come very early in the development of their product? 15:26 Because obviously with hardware it's a very different kind of cycle than it is for software. 15:30 Um, and what does that all look like, uh, just for, you know, I'm sure a number of audience members out there are interested in kind of understanding how that works. So, so it's a pretty broad spectrum. 15:40 So certainly, uh, we have a mixture of us approaching customers and customers approaching us. 15:46 We, we, e- early on, we were pretty selective just to get good feedback on what we were building and to focus our own efforts as a relatively small company to make sure that we could satisfy our customers' needs. 15:59 It's been a pleasant surprise w- uh, coming out of stealth, you know, we were, uh, s- kept ourselves secret mostly until this past March, and even though we announced ourselves during a pandemic, 16:11 it, it has been shocking in a good way to see how many folks have come out of the woodwork, and we're getting ridiculous amounts of interest now from all over the world, um, and are in fact having to juggle to make sure we can manage the level of interest that we're, that we're seeing. 16:27 Um, to, to your question about the nature of the engagement, what typically happens 16:32 is that, uh, either there's a current product that our customer wants to upgrade with Ergo, or they're looking at some next generation product they wanna build where they see the value proposition of having an edge gadget be intelligent, whether that's battery powered, and we're extremely low power, whether they need sophistication of inference, being able to do recognition of objects under difficult circumstances. 16:58 Uh, 16:59 th- those are the, the, the basis of the engagement, and then we sit with their team, their hardware people, their software people, and help them design what their product will look like, uh, given all the horsepower that, that Ergo can bring to them. 17:14 What applications do they need? How should they hook this up with the hardware that they have, et cetera. 17:20 Now, uh, let me ask you, as you talk to more and more customers, have you seen any kind of change in demand from the market due to being in the pandemic? 17:29 I know we've had a number of guests on that are in the hardware space, a number of guests who are in the, um, smart home space, and because people are spending more time in their homes, there's been a higher demand for, for products, smart products to be placed inside of their home. 17:44 Um, and security is often a big concern, and I'd be curious just to get your take from your point of view on how you've seen COVID kind of affect your business and the industry that you, you play in, um, over the last, you know, six to s- six or so months. 18:02 I, I, I agree with both of the points you're making. We've seen it too. 18:05 One is that we are definitely seeing a broader interest, I think because of the pandemic and people spending time in their homes, and the second point that people are more sensitive about privity, privacy and security even than they were, uh, six months ago. 18:22 So o- on the first, that everybody's in video conferencing so much, uh, that, so people are using, uh, video and so on, so havingUh, object recognition, having voice recognition, having the, the, uh, video conferencing track where my eyes are so it looks like I'm looking at you even if I'm not. 18:44 Uh, having your grandmother, who doesn't quite know how to use an iPad, still be able to do video conferencing by having the i- by, by having the, the, the tablet be a little bit smarter about the, uh, conversation itself, are things we're seeing. 18:59 And on the privacy and security side, uh, sorry, on the privacy and security side, it's people are increasingly sensitive to the hacking that can happen in IoT and, and really wanna keep their data private, and we can help do that. 19:13 So that's a very good point, and this is something that many, many months ago, I forget exactly who the guest was, but we spoke about this in back-to-back episodes about the privacy of smart home devices and where the data is going and who's controlling the data. 19:28 And I know you've mentioned this m- a couple times already in our conversation here, but this particularly relates to the hardware devices that people are buying, um, and not really understanding who is actually manufacturing and supporting these devices. 19:44 And oftentimes they're white label devices that were originally made in China, and the data is being sent back, like you said, to, you know, to a cloud that is... 19:55 Often you don't know who's controlling it, and in a lot of these t- situations is a Chinese company or organization which right now, um, is even more of a hot topic, uh, about where the data is when you think of TikTok and these apps out there that, you know, is, are started or, or, you know, launched in China and people are concerned about the data. 20:16 So how do you all approach that type of conversation when you're working with either clients or you're talking to just educating the mark- or talk- educating the market in general, uh, about the smart home kind of device buying process, and why your solution enables a better, um, kind of touch point or a better, um, offering from a benefit standpoint for customers because you c- you're helping solve that security issue that often people are very kind of... 20:47 I guess they're n- they're not, um, they don't know it even exists, or they don't pay too much mind to it. So, uh, uh, thanks very much for asking about that. I think this is a super important topic. 20:58 And frankly, whether the person who's watching the inside of my house is in Illinois or i- or i- or in, uh, Shanghai, [laughs] i- i- i- it, both of them are creepy from where I sit. 21:10 Um, i- i- we have, uh, uh, addressed this issue of privacy and security in two important ways. First, we don't have your raw data leave our chip at all. So, so the sensor is... 21:25 Uh, the camera or the microphone are receiving raw data. It goes into our chip. Our chip provides inferences, that is what's being said, who is there, and so on, without providing the raw data anywhere else. 21:38 So it's never going to somebody's cloud, uh, uh, where it's not controlled. The second thing that we're doing for security besides that is that we're encrypting absolutely everything. So we encrypt... 21:50 All of the code that runs on our chip is encrypted. 21:53 The inference engines, everything is encrypted, and we're doing that as a way of at least taking a serious whack at mitigating hacking by, uh, foreign agents or, or, or malicious, uh, characters. 22:07 Um, w- I can say that from a customer point of view, we get a great deal of resonance for this that, um, they're hearing from their customers just as we're hearing from ours the importance of privacy, the importance of security, and it just hasn't been technically possible before now to do all that processing at the edge, and part of why they're excited about us is we're the first company that can actually do that. 22:30 Wow. Okay. Very good. 22:32 I've, I've, I've always been curious kind of, uh, to expand on that and figure out what a solution would be, um, because most people, you know, don't pay too much attention to it, um, or don't really know enough about it to be educated to make the right decisions, and I think any company out there that's trying to solve that problem or has solved that problem is, is, is fantastic. 22:56 Now, one question I did wanna ask is, since [clears throat] you're keeping the data at the edge and, you know, it's all being processed there, are th- is there any, like... This might be a loaded question. 23:05 Is there any downside to, to taking that approach versus having the c- the data go to the cloud and, and be handled there, like from a cost perspective? 23:14 Is there a, like, you know, basically is there any trade-off to doing that? Well, I guess I'd say two things. First, it's not like we're trying to replace the cloud. We complement the cloud. 23:25 I think if I wanna find videos that have you in them, that's a good question to ask the cloud because the cloud has access to all the history of the world. 23:34 But if I wanna change this set of pixels, this picture, to say, "Is it Ryan or is it Steve?" that's something that should be at the edge. 23:41 So I think we're, we're really making the same kind of separation that nature has made. You still have eyes even though you have a brain, and your eyes and your brain aren't the same thing. 23:52 So I think the same should be true, uh, in, in electronics where you have sense organs out there that are smart, eyes and ears and so on, but there's also a brain that has memories of what's happened to you over the last 20 years. 24:06 So I think that's the appropriate, the appropriate separation. Okay. That makes a lot of sense. 24:11 Um, [clears throat] so I wanted to ask you about the challenges you kind of encountered throughout this process of developing Ergo and breaking this company off from, uh, the parent company into its own, own individual business. 24:22 Um, uh, so it seems like you guys have s- are obviously redefining the machine learning framework for edge devices. Uh, can you talk a little bit about the challenges that you encountered doing that, and maybe how... 24:34 And you've already expanded on this, the second part of the question already, but maybe if there's anything you can think to add to it. 24:40 But how Ergo has differedFrom alternative edge, um, inference processors that are out there on the market. Um, so got a two-part question, but, uh, love to kind of get your take there. 24:53 Well, this is a very ambitious undertaking. Uh, one of the things as you look at historical efforts to build inference gadgets 25:02 is they're typically either started by people who know a lot about hardware but not a lot about machine learning, or people who know a lot about machine learning but not a lot about hardware. 25:12 And as a result, there are sort of cracks at the interfaces. 25:15 We have deliberately started our team where everything is sort of integrated, where we are designing the machine learning strategy and the hardware simultaneously, and we're doing that on purpose. 25:27 If you really wanna get something efficient that, that, that, that really sings, um, you really have to think of the system as a system. 25:35 And while that's made the project a lot of fun, [laughs] it's also made it very ambitious, where we have to have a collection of really gifted, uh, uh, chip architect type people and machine learning type, uh, people all arguing with each other in figuratively the same room, uh, to, to, to, to get to an overall hardware software system that really works as a unit. 25:58 So that, that's certainly been challenging, but that's also been part of the fun. Gotcha. Okay, fantastic. 26:03 And the last question I have before we wrap up here is, what is your just high level kind of prediction for the future of, let's say, smart home devices now that, you know, a solution like Ergo exists? 26:16 Like, where do you see the market going, um, and what are you most excited for? Oh, I, I... Our ambitions are significant here. 26:25 The reason I called the company Perceive is because we wanna upgrade the whole idea of sensing into perceiving. 26:32 I see it as a historical accident that we have sensors that are sort of dumb in the sense they just hand you pixels or hand you audio but have no idea what they're looking at or what they're listening to. 26:44 The possibility that wherever you would have a sensor, and there are twenty-five billion sensors right now, there will be a trillion in 2025, that every one of those sensors can kind of understand what's going on around it, I think can make every gadget we interact with have a more natural interface, be something you can talk to and interact with in a way that has simply not been possible before Ergo. 27:09 So I, I'd like to see Ergo be everywhere that you'd have sensors, and I think that's what should happen over the next several years, Ergo and its kin. Yeah, that's awesome. That's fantastic. 27:19 Yeah, I'm super excited for it. 27:20 You know, we don't, [clears throat] we don't cover too much on the consumer side, um, but we can, we try to talk more about the technology that's behind a lot of these, um, consumer applications and c- and devices and products are out there. 27:34 So having somebody like you on the show today has been fantastic to kind of educate from a very different angle, um, that I think is, uh, you know, is an enough... 27:44 Do you ha- You guys are basically approaching the problem from a, from a unique side of things to really 27:51 handle a lot of the concerns at least I've had as a consumer, and I've seen other people voice, like the security issues, the, the data issues, and all those things. 28:00 So, so what you guys have going on is absolutely fantastic, and I really appreciate you taking the time to share these insights today. 28:07 One thing I do wanna ask you before we wrap up is, is there anything happening on the Perceive side of things in the next couple of months that our audience should be on the lookout for? 28:16 I know you guys mentioned coming out of stealth mode, and you, you have, you know, c- customers that you're gonna soon be able to publicly reference. 28:23 Um, but just anything that we should be paying attention to or be excited to kind of see coming out of your guys' shop. 28:30 I, I think in the short term, it's mostly that you'll see customer announcements of what products are coming and so on. 28:35 But, uh, I, I guess I'll make the comment that, uh, this is only our first generation of products, and wait till you see what's next. We have a whole roadmap of cool things coming down the pipe. 28:47 Well, when that happens, we expect you to be back on the show to tell us all about it. [laughs] I look forward to it. Thanks very much for this opportunity. Absolutely. 28:56 Last thing I wanted to do before we jump off here is, um, audience out, if anybody in the audience out there is, uh, interested in learning more or contacting you, following up, so forth, what's the best way to do that? 29:07 I would start with our website. Okay. Which is at perceive.io, and you can find out more about us and also reach us through that. Fantastic. All right. 29:14 We'll make sure we link all that up and all the distribution and content that we write around this episode. Um, but again, uh, Steve, thanks again for your time. 29:21 Uh, really appreciate it and look forward to having you on the show again in the future to announce kind of anything new going on at Perceive, and we look forward to kinda seeing, seeing what you guys, um, develop and bring to the market. 29:33 Thanks a lot, Ryan. I really appreciate it. Absolutely. Thanks again for listening to that episode of the IoT For All Podcast. Hope you found a lot of value in it. 29:39 If you did, we'd truly appreciate it if you would leave a rating or review on whichever platform you're listening to us on. 29:43 Also, feel free to subscribe to our channel, uh, podcast channel directly, or feel free to subscribe to our newsletter or social media so that you always get the latest episodes as soon as they become available. 29:54 But on that, thanks again for listening, and we'll see you next time.