In episode 19 of the podcast, the topic is: Machine Learning in Manufacturing. Our guest is Michael Zolotov, CTO & co-founder at Razor Labs.
In this conversation, we talk about where we are with machine learning and AI for manufacturing. What are the main techniques? What is possible now? What will be possible soon?
After listening to this episode, check out Razor Labs: http://www.razor-labs.com/ as well as Michael Zolotov's profile on social media: https://www.linkedin.com/in/michael-zolotov-33a2b26b/
You may want to also be aware of the 'Israel meets New England' smart manufacturing event on June 9 and its organizers, the Israeli Trade Mission and Amhub New England:
Trond's takeaway: Machine learning is definitely entering manufacturing over the next few years. Already, interesting experiments are underway to do simpler things such as prevent future downtime using sensor data already being captured by advanced machinery. Pure machine optimization can only get us so far, though. The real potential lies in complex business process optimization and simplification with augmented frontline operations. Technology plays a part, but clever workers, operators, and engineers will have to make intelligent use of the technologies available, they cannot just blindly implement. For that, we need reskilling--always learn.
Thanks for listening. If you liked the show, subscribe at Augmentedpodcast.co or in your preferred podcast player, and rate us with five stars. If you liked this episode, you might also like episode 27, Industry 4.0 Tools, episode 13, Get Manufacturing Superpowers, and episode 14, Bottom up and Deep Digitization of Operations.
Augmented--upskilling the workforce for industry 4.0 frontline operations.
#19 Machine Learning in Manufacturing_Michael-Zolotov
[00:00:00] Trond Arne Undheim, host: [00:00:00] Augmented reveals the stories behind the new era of industrial operations, where technology will restore the agility of frontline workers. In episode 19 of the podcast, the topic is machine learning and manufacturing. Our guest is Michael Zola, Tav CTO, and co-founder at razor labs in this conversation, we talk about.
[00:00:24] Where we are with machine learning and AI for manufacturing. What are the main techniques what's possible now? What will be possible? Soon? Augmented is a podcast for leaders hosted by futurists Trond Arne Undheim, presented by Tulip.co the frontline operations platform and associated with mfg.works the manufacturing upskilling community launched at the world economic forum. Each episode dives, deep into a contemporary topic of concern across the industry and airs at 9:00 AM. U [00:01:00] S Eastern time every Wednesday, Augmented the industry 4.0 podcast. Michael, how are you today?
[00:01:08] Michael Zolotov: [00:01:08] Hi, Trond. Great being here.
[00:01:10] Thank you.
[00:01:11] Trond Arne Undheim, host: [00:01:11] Yeah, sure. This is exciting we'll talk about machine learning and manufacturing which quite a bit about, look, I'm excited. You're an engineer from from Tel Aviv and then you went ahead and obviously did a military thing. Many of you do in Israel and I'm super excited about that.
[00:01:28] So let's chat about it for a second and then, and then obviously now founding a startup, which we'll really talk about, but what is this talk, to your pro program, the next R and D leaders in the Israeli military. What's, what is the public version of what that is all about?
[00:01:43] Michael Zolotov: [00:01:43] So basically, as military service is a mandatory service here in Israel, every boy and girl in the age of 18 is listening to the army. And we're trying to to find out technical solutions to their problems, to the challenges, the threats we're facing. Tell PRT is a [00:02:00] program where essentially they pick 50 of the brightest minds every year.
[00:02:05] To be the ones they call the next R and D leaders of the army and basically one of the key challenges that the army faces is that there's such a huge flood of information. And the goal of some of these leaders after they finished their program, it's a three-year program is to Use artificial intelligence and specifically deep learning to be able to fuse all these information sources to basically whether it's detect threats get insights, intelligence, and so on.
[00:02:37]It puts together both academic studies. Physics and computer science together with military studies and after that you are placed in one of the of the critical points of the idea of the Israeli defense forces and specifically what I've dealt with is fusing together. Radar signals into to be able to detect threats on the country of Israel.
[00:03:00] [00:03:00] Trond Arne Undheim, host: [00:03:00] It's fascinating because you have to get the skills somehow and in this case, you get to apply some of your techniques from university in a practical setting. But anyway let's move on to industry 4.0 and how you have started to apply it. Fast forward a few years, and you're, you're obviously out of the military forces and you get together with some of your buddies and you form a startup.
[00:03:23] How, what was it natural for you to move into? Basically the manufacturing side. Is that something that was on your mind or were you just interested in applying your machine learning skills?
[00:03:35]Michael Zolotov: [00:03:35] So it's more of the second option. When we were discharged from the army, we, were in the we were in the field really from the Dawn of deep learning and we were looking for the, really the the sector or the market with the highest impact that this technology can make and after inspecting several markets, we found industry 4.0 as the place to perform a gigantic impact on the market [00:04:00] with this technology. So that's what drove us.
[00:04:04] Trond Arne Undheim, host: [00:04:04] So industry 4.0 means a lot of things in the US they typically go for this term smart manufacturing, at least when it's applied directly to Manufacturing, because it's seems to be a very, it's like you're very inclusive and ambitious term also.
[00:04:20] What does it mean to you and specifically when you said you can make a, an enormous impact with deep learning what do you have in mind specifically? Give us a little deep dive into deep learning and Manufacturing.
[00:04:35]Michael Zolotov: [00:04:35] Okay. Would you like me to start with a really short intro of what's deep learning and how it differentiates from AI machine?
[00:04:41] And there's so many buzzwords out there.
[00:04:43] Trond Arne Undheim, host: [00:04:43] Yes, please, please.
[00:04:44] Michael Zolotov: [00:04:44] Okay. So basically AI is really a buzzword today, but it existings the fifties. It's not a new thing and it was actually built to solve complex problems that it was only thought that humans can solve them. That's how it started in [00:05:00] practice and obviously in the beginning it was made of simple rules that really limited the capability of AI to solve complex problems. And starting from the nineties, we had machine learning that came up as a brand new technology. And the thing with machine learning is for the first time, the computer can actually learn from experience, but the human needs to define the exact features or parameters that are important.
[00:05:26] So let's take the most simple example. Let's say I want to differentiate between a dog and a cat. So the features that I'll choose is let's say the color of the, for the distance between the eyes and so on. And the machine learning algorithm, we find the optimal weights on the importance of a feature of these features to perform this differentiation.
[00:05:47] The problem is that because you're are limited to the features that the person. People obviously extract. You're very, also limited in the complexity of the problems that you can choose. And you could see [00:06:00] that we really got to a very two glass ceiling in many fields and that's where deep learning is a totally new thing.
[00:06:08] Deep learning for the first time allowed us to break this glass ceiling and this is what allowed, Siri to understand what you're saying. It allowed Facebook to detect faces in, in photos. It allowed the autonomous cars and so on and deep learning centrally finds itself the features that are important itself.
[00:06:26] It can choose millions and millions of parameters. Many times invisible to the naked eye and can actually make sense of them based solely on examples. So in this case, you just give 1000 examples, cats, 1000 example of dogs and it by itself understand how to do that and we essentially simulate how we as people, how our brains work in a computer.
[00:06:49] So that's the revolution that affects really every part of our lives in so many applications and what we do in Reza labs is we bring [00:07:00] this technology, this revolution to the industry 4.0 where essentially the way that we see it is take these machines that up until now are not aware of their environment. They're just machines with a very specific programming that performed their task they're not aware of what's happening and using these very sophisticated tools of deep learning. We can make them adapt to the environment, to this constantly changing environment and actually produce more with less materials and be able to optimize whatever KPI is important to the client. It can be maximize throughput. It can be lower emissions. It can be more higher energy efficiency, or it can be a combination of all of these factors together.
[00:07:48] Trond Arne Undheim, host: [00:07:48] And what are the inputs that you're using? Do you just plug into the machines wherever they are and whatever their sensors are, because in the dog and cat example, you're assuming a vision [00:08:00] input of some sort like an image, but I'm assuming for many of these machines, it's not vision.
[00:08:06] But although vision can be one of them, with cameras or other types of computer vision, but it is other types of sensors. What are the most typical sensors that you'll find on industrial equipment in factories these days? .
[00:08:20] Michael Zolotov: [00:08:20] So that's an amazing question essentially what we do, every machine or most of the machines come out of the box with some sensors on them. And the very simple reason why the sensor are there are in order to make the PLC. The PLC is the logic that drives the sensors the machine, sorry basically make the PLC drag the machine. I'll give the very basic answer. The very basic example.
[00:08:44] Let's take an autopilot on planes in oil for the autopilot to work it uses sensors on the plane, such as the wind speeds, such as the altitude of the air of the planes, such as the power of the engines. And so on, do this, all these sensors in order to drive or do to [00:09:00] automatically fly the plane. So each of these gigantic machines already come with sensors that are used by the PLC in order to drive the machine or to operate the machine and we leverage these existing sensors in order to fuse all of them or analyze all of them using deep learning. So the sensors are already there. They've been there for many years, gathered terabytes and terabytes of data that's stored and not leveraged today and that's when we come into place.
[00:09:28] Trond Arne Undheim, host: [00:09:28] Got it and deep learning is no, it is a paradigm. How wedded are you to deep learning? Because there are also other paradigms out there. Obviously it was a big kind of breakup moment a few years ago, with with some deep learning data sets being released. But what are some of the other techniques that you're using specifically?
[00:09:49] Michael Zolotov: [00:09:49] So we actually we use solely deep learning and we focus specifically on deep learning because it allows us to break many barriers that are out there in the market. [00:10:00] So essentially most of our competitors use the more the more classic or the older, the old fashioned machine learning techniques, which might be less complex algorithmically but eventually limit both the accuracy that you can achieve and the generic of the algorithm. So you can solve less problems and with less accuracy and therefore we only lived ourselves to, to deep learning, which do not have these limitations.
[00:10:28]Trond Arne Undheim, host: [00:10:28] And I understand within the learning, which is what's called a neural network approach by the analogy to the brain, which I find a little questionable.
[00:10:37] But anyway I know you use reinforcement learning algorithms specifically. Can you just explain that in, a, in essence, what is reinforcement referred to. .
[00:10:47]Michael Zolotov: [00:10:47] An amazing question. Let's say you want to optimize a machine and let's say that your KPI is you're happy with your throughput, but you want a constant throughput.
[00:10:56] You want it with minimal variability. So [00:11:00] essentially if you, in the real world, what you wanted to do is you would have You're you have the parameters that you can change. For example, let's take mil, you can change the speed of the mill. You can change the weight of the mill.
[00:11:13] There are many parameters you can change. Obviously, the composition of the materials going into the mill and obviously meals are used in so many manufacturing plants. And in this case, what you would have wanted to do is you would want to perform some action and see how the machine reacts. Would it perform another action and other actions?
[00:11:32] Obviously that's not something that you can perform in real life because. I mean that's not something that that you should do. So what you do is you take this smell and you essentially create a digital twin of this smell in a computer, and it is composed of two parts. The first part is essentially a simulator.
[00:11:49] The simulator does exactly that if you are again, in a computer perform some action. The simulator tells you what would have happened if you would have done it in real life. [00:12:00] So this is the first component that simulates how the machine works. And then you have the second component, which is the agent and the agent performs billions and billions of actions.
[00:12:09] Everything digitally, many times the amount of actions accumulate to hundreds of years of the real world, if it would have been performed in the real world and it performs so many actions and see so many reactions that it is able to create, what's called the policy. Or the optimal policy that would maximize the KPI that you chose.
[00:12:31] So in this case you want minimum variability of the throughput and therefore you're essentially this agent now has, because it has seen hundreds of years of mil of milling simulations. It essentially has more experience than any meal operator out there in the whole world. And then you can take this agent and you can deploy it in the real world to give recommendations to the human, actually driving the [00:13:00] meal, actually operating the mill.
[00:13:01] So this is called reinforcement technology reinforcement learning. Where through reinforcement through action reaction. You're able to learn a policy that maximizes the KPIs that you choose. And many cases, different clients choose different KPIs. One client can say, I want the same throughput, but with minimal emissions, another client can say, I want the maximum throughput given the constant or minimal amount of malfunctions in the system.
[00:13:29] So many clients have different KPIs, but this agent can optimize any KPI as long as you can define it.
[00:13:36] Trond Arne Undheim, host: [00:13:36] So it's like experience on steroids. It would be like you had a factory with thousands of people and you could run it for a thousand years or a hundred years and then just look at all the experience that you would have had.
[00:13:48]It's fascinating. So here's my question. If you are a factory manager or a plant manager of some sort, and you're interested in this topic. We'll get into what razor labs specifically [00:14:00] does in a second, but if you just want to understand the field of deep learning, applied to Manufacturing, or even just various machine learning approaches, what is your best guess as to how they should do that?
[00:14:12] Are there courses available? Can you just go on YouTube and look it up, or should they take six months or a year and study it in the university? What should one do these days to stay up with all of these techniques because not everybody has three years and can be chosen as the top 50 in the Israeli military.
[00:14:29] So we have to have some other options on the table.
[00:14:32] Michael Zolotov: [00:14:32] Of course. So it obviously depends on who the stakeholder is. Obviously a CTO, a chief innovation chief data officer in a company already, they mostly come from engineering backgrounds and have this knowledge many times if the knowledge is does not include specifically deep learning, but does include machine learning.
[00:14:51] Many times the gaps are not large and whether it's through a universities or just courses on Coursera, you then Demi and many others are open [00:15:00] Stanford courses. The gaps are not large for business stakeholders, which is I think the most, the more interesting case to discuss I think that what's most important is understanding the applications of AI and deep learning.
[00:15:15] It's less important to understand exactly how they work. It is important to know what we need in an organization or what the organization needs in order to be able to assimilate. AI. And what are the limitations of AI? Many of you have these you have these crazy assumptions that if you assimilate the AI, you don't need any people anymore and it will just replace all of them.
[00:15:37] So it's very important, I think, to understand the applications and the mutations. It's less important for them to concentrate on exactly how we structure these annual networks.
[00:15:48] Trond Arne Undheim, host: [00:15:48] That may be true, but some of my not transparency, right? I wouldn't want to be running a factory and had no idea how the data that I'm basing my decisions on, how it's generated.
[00:15:58] So there's some balance there [00:16:00] between, the level of knowledge, let's switch to, to raise our labs. You have an interesting origin story. You met a gold mining company on a trip to Israel and and then things started to happen.
[00:16:14] Michael Zolotov: [00:16:14] Yeah. So it was a, quite an, a, an amazing story. We had a NuQuest, which is the world's third, a gold mining company coming with a delegation to Israel visiting our offices.
[00:16:27] We were much smaller back then and they presented us with some of the, with some of their challenges some of the issues that they were facing. And obviously, we have no mining in Israel have no natural resources recently they did discover guests here, but no methods and no, no gold. And we presented them the technology and we presented similar use cases, but applied to other sectors and they actually chose us to be to be their partner in their jaw journey of assimilating AI. And it's [00:17:00] an amazing collaboration that that lasts and still lasts for many years where essentially optimizing their entire operations, bringing autonomy and effectiveness to their operation through their simulation of AI.
[00:17:13] And it actually came to the point where their CIO came to the Australian papers and published an entire article about that, which obviously gave us a very substantial boost in the Australian market.
[00:17:25] Trond Arne Undheim, host: [00:17:25] And you bootstrapped the startup up to 60 employees, and then you did something non-traditional for, even for an Israeli tech startup, because you, you did an IPO and you're now listed on the Israeli stock exchange as the first tech, arguably the first, at least for the first kind of machine learning company.
[00:17:44] And first AI company on the Israeli stock exchange, which is different from most the typical Israeli startup route would be, you go to the U S and try to make it big and set up a subsidiary. And then you scream and shout until you, maybe you get listed on NASDAQ or something, but why [00:18:00] did you go this route?
[00:18:02] Michael Zolotov: [00:18:02] Instead of going to VCs, you mean
[00:18:04] Trond Arne Undheim, host: [00:18:04] instead of going to VCs first, and then, obviously what you do after this is another story but just this decision to go on the Israeli stock exchange and not go to VCs for the first move.
[00:18:15]Michael Zolotov: [00:18:15] Starting from literally the inception of the company, it was first of all, very important for us to be independent, having the independent say and have only are the three partners in the board and having that in addition to the fact that we wanted a very A very clear cut product market fit. And we knew that if our product was good, we would we would be able to bootstrap the company. And that's exactly what we did. So taking the two of them together, actually, and, obviously there is a very high demand for artificial intelligence solutions for these markets, because it's really easy to to convert. Supreme technology to just pure business value to pure dollars of whether it's additional [00:19:00] throughput of whether it's less downtime. It's very simple. And through these demands on product, we're able to grow to a roughly 60 team members and the goal here is not is not an exit strategy. We wanted to establish, a serious and prospering enterprise and that's why we went to an IPO that we felt was the best way for us. And we started from the Tel Aviv stock exchange. But our next goal is the NASDAQ in two years. That's where we were aiming.
[00:19:31] Trond Arne Undheim, host: [00:19:31] That's ambitious. All of you guys are ambitious guys. Tell it, tell me a little bit about some of the clients that you have and what you do for them.
[00:19:39] So I understand you're involved in it. We've talked about natural resources, right? So that's how it started. And then we had talked a little bit about manufacturing generally and how you are taking on that market. And then the third market is utilities. What do you do for these various segments specifically?
[00:19:56] Is it machine monitoring, like we've talked about [00:20:00] before, or are there other specific things that you can accomplish for these clients?
[00:20:04]Michael Zolotov: [00:20:04] So there are several you might say use cases, the most frequent use cases. The use case that we get from our clients are prevention of future downtime, because that's a clear cut on. You prevent this downtime, it's additional throughput that's being created. That's the majority of the use cases that we get. In addition to that eventually we have the O E the overall equipment effectiveness measure and. What's important for the business stakeholders is to maximize this Zoe and they're willing, and they understand that AI is a very powerful tool to perform this maximization from our research, we know that an average OEM on the Western country is roughly 65% in our factory. So there's a lot of room of improvement. And many times starting from the downtime of the machine is a really is a really good place to start because we really get to the 80 20 of the value that you can get. So we're speaking about prevention of downtime. We're speaking [00:21:00] about condition-based monitoring, we're speaking about indication in real-time improper maintenance done to the machine.
[00:21:06] So it's literally a system that's in real-time ingesting all these sensors, many times tens of thousands of sensors and very intuitively giving you these insights, whether it's in a dashboard on your smartphone other clients ask for SMS messages, any any way that they want prevention of quality issues and so on and obviously machine optimization for some given a KPI as we discussed previously.
[00:21:32]Trond Arne Undheim, host: [00:21:32] There are there are a lot of different targets that you could have if you are a company in any of these segments, and everybody wants oper operational efficiencies and even small efficiencies can make a difference.
[00:21:42] What do you think about I guess the current situation. So you shared with me a little bit earlier, and this is not the bad mouth clients, but it's just generally in the market, there is a little bit of bewilderment around what data can do and arguably, [00:22:00] most organizations, obviously not your clients, because they have started to take this on board, but, We have known about big data or any data on the importance they could have in, on a business for a while arguably, why is it that the operations practice of many factories and manufacturers, why have they been slower? Arguably to take on and start using these kinds of metrics and maybe even just are actually not even exploring this, thinking that it is, it's something we will get to. Why would you say that maybe perhaps smaller factories especially have been pretty slow, whether it is to take in advanced machines. So even experiment with robots, like all of this stuff it's very much talked about. It's very much out there as the newest biggest thing and everybody claims that they are looking, but comparatively few [00:23:00] have actually implemented them at scale, at least pre pandemic, in their operations. Why is that?
[00:23:06]Michael Zolotov: [00:23:06] First of all when we first got into the sector, we obviously wanted to hear the thoughts of executive in this sector about assimilating AI and the first and that's, it's really in my memory that a survey made by Accenture says that three out of four C-suite executives believe that if they don't embrace AI and then don't use AI in their operations in the next five years they are risking their business.
[00:23:33] So while there are executives, out there, or there are factories or plants out there that do not leverage AI. Today you have more and more, and with very high pace and very high increase of numbers of plans that are active in factories, actively looking for AI solutions to assimilate in order to maximize OEMs.
[00:23:56] I think that There are two main factors here. The [00:24:00] first factor is the data up until recently or up until several years ago, most companies or most factories simply did not store the data that was gathered. So you have the sensors were used for the PLC, but nobody's started the sensors and the installation of historians for many clients, for many factors has been there only for the for the past few years and obviously with no data, you cannot leverage AI. And the second component is, a vicious cycle where you store data. The data originally is not of such a high quality because it's not a high quality because it's of low quality. It's very hard to actually leverage it for some business value.
[00:24:42] If you don't leverage for business value, you have no incentive of improving the quality of the data because it takes effort. And so on, it goes into a vicious cycle and machine learning tools are simply not powerful enough in order to leverage this data with so high discrepancies. [00:25:00] And only when you get to deep learning, which is really technology of the past few years there, it has the complexity in the models that's able to overcome these discrepancies.
[00:25:10] And that's also one of the key reasons why we went in that direction to today. When we take data from a factory, from a manufacturer, the first thing that the manufacturer says is. We tried working with so many companies and they gave up because the data was not good enough and we always smile because we hear it so many times.
[00:25:29] And the end, you actually get that with machine learning. It has so many discrepancies, sensors are being replaced without any warning. Parts of the entire machine are being replaced. The entire statistics of the data just changes and the model is not aware of that. And you've got to have this super this high complexity of deep learning to be able to not only perform this optimization or prediction or whatever, but also track these changes whether because of a maintenance or because of some other [00:26:00] changing environment and be able to incorporate them in the model so we call it evolving AI that adapt itself to the changing environment.
[00:26:09] Trond Arne Undheim, host: [00:26:09] I wanted to talk about just one very specific challenge that I've at least heard others speak about. If you think about, deep learning, which we've talked about now it's comes from the it side. It comes from the software side now operations plant operations and generally the field of operations management is more of an engineering culture that traditionally doesn't really rely on software. And it's a different kind of a vibe it's a different Culture perhaps, but from you, your side, you are coming at it from the software side. What were the learning challenges you faced in trying to understand what and Manufacturing business is. And because because they're not irrational challenges that these people have. They're actually trying to operate a factory to try and get a shop floor to work seems very different [00:27:00] from the mind of a bright it person, anywhere in your organization. It's just trying to optimize an algorithm there, there a little bit. There's a little bit more complexity there. Real life complexity.
[00:27:10] Michael Zolotov: [00:27:10] You're totally correct. Eventually that's exactly the difference between OT and it, the very the very much the vast majority of the personnel in a mining facility, in the mining site or in a manufacturing plant or a factory. They come in the morning to work and their goal is to make sure that the machines work today and tomorrow and next week. And they deal with the day-to-day challenges of the operations. And therefore, I think that one of the key critical things of a factory that wants to assimilate AI is to have the innovation department that can basically, it's a bridge between startups like us. And the operations, because you cannot expect from an operation, whether it's whether it's if you predict functions, you cannot expect from a [00:28:00] maintenance. Intense superintendent or improvements of superintendent to leave all their work and start tweaking AI algorithms. So that's the goal of the innovation department. And they're look for obviously many years into the future and they established the pilots. They establish the proof of values and from there, you can go further into the actual assimilation in the OT operations in the day to day life of the personnel there.
[00:28:29] Trond Arne Undheim, host: [00:28:29] How difficult is your software to install, right? There's the big movement now in it to go towards more what's called low code, or even no code environments where people MOTIE or really any engineering background.
[00:28:42] And even just someone who is used to Excel sheets can start to engage on these issues. Where are we? I could just imagine that for deep learning, the challenge is quite significant because you need to build not only just the basic software skill platforms, to make them [00:29:00] usable, but you actually have to build the algorithmic level and some of the inputs and outputs and connect it to the sensor.
[00:29:06]There's a bunch of challenges before that can be made simple. Where are we? With this challenge?
[00:29:12] Michael Zolotov: [00:29:12] Are you familiar with wix.company? So wigs are essentially cordless solution for websites. And today there are many companies that are essentially like the weeks of of AI, essentially with zero lines of code, you can build an AI model, whether it's machine learning and deep learning or deep learning and the thing that is, do you have a trade off? Obviously the advantage is it's no code. The disadvantage is that you can only build very simple applications, right? For example, differentiate between a dog and a cat. That's something that today you can build with almost no code. Fusing together, tens of thousands of five tens or thousands of five tags in and dealing with all these data discrepancies is something that you cannot do with no code today. I don't think you can even do it with no code [00:30:00] in the next many years, because the complexity there requires you to be a really, a very high expert in deep learning. And what we put a very high emphasis on, and that's where we invested years of R and D is to be able to come with models that are already pre-trained on a very vast library of components.
[00:30:25] And to have this adaptability to the changing environment. So for example, if I learned on a motto or a drive of this of GE, and now I need to adapt it to Siemens, it does that automatically without me being involved. So that's really cutting edge technology. And from the client's side, it allows us to very swiftly and very quickly deploy the models and with really minimal Involvement than had that from the client because the AI does that by itself, but it took many years of firing the together.
[00:30:54] Trond Arne Undheim, host: [00:30:54] If you think about the coming year, what are some of the industry developments that you might be excited [00:31:00] about it? If you were looking at it from the perspective of where your field is moving or indeed, where industry 4.0 is moving how slowly does this move and what should we be expecting in a, very short term?
[00:31:13]What are the things that are happening in the industry? From your point of view right now.
[00:31:19]Michael Zolotov: [00:31:19] You have this pyramid of AI adoption. So from our experience, more, most of the clients are really in the beginning of the heir journey. Some of them have never tried AI at all.
[00:31:29] Some of them are really playing with the first with the first POC pilots of AI. So we expect more companies to have AI embedded within their operations. And we need to understand it obviously there's the technological part, but we have the not less important change management part because AI really changes the way that these operations work.
[00:31:54] Suddenly a person instead of, doing the tedious work on the gets insights from [00:32:00] the AI in real-time and now focuses more on the creative side of things, or only on approving the decisions and not literally performing the optimizations along. These change management of assimilating AI in the operations.
[00:32:14] I think we'll see more and more of that when all these POC is become deployments and all these deployments will become the part of the day-to-day operations and life of the OT and not the idea of the OT section. From the technology perspective today, most of the use cases and most of the, yeah, most of the use cases are concentrated on prediction, predicting something, whether it's predicting throughput, variability, predicting energy efficiency, predicting malfunctions that's.
[00:32:45]And the reason for that is because that's more basic and you've got to start with something. I think that in the next few years, we'll see more shifts towards optimization when a client. We'll define an overall or a holistic [00:33:00] KPI, whether it's minimal amount of new function, maximum throughput, minimum variability, we don't pollution so on and in AI we'll holistically, not only look at the machine, but in the entire process and optimize the entire production line. In order to meet these KPIs maximum is the KPIs. So that's where I think things are heading. I would have, I would add one more thing. I think that one of the also critical things in deep learning is having what's called external mobility so you mentioned it briefly. It's very important for the stakeholders whether it's onsite or the business, the colors to understand the why and the, see why AI performed some decision, why it performed some prediction or some optimization. It cannot be a black box because eventually. The impact that it has, whether you need to send teams thousands of miles to do some location.
[00:33:52] When you, whether you need to change your operations, it cannot be based on a prediction that you don't understand. And [00:34:00] that's something that we put a lot of effort and a lot of R and D into that, to make sure that these black boxes are open and understandable.
[00:34:08]Trond Arne Undheim, host: [00:34:08] It, it seems to be. Exciting times to be in manufacturing.
[00:34:12] How do you explain that? People generally don't quite see this. If you think about the skill shortage in manufacturing, the talent gap, the fact that so many young people. Don't make the choice you made and, create a startup in manufacturing or at least targeting manufacturing there. How do you explain this discrepancy is so all of these advanced technologies industry 4.0 like robotics sensors, right?
[00:34:36] Control systems and indeed deep learning, like the hottest thing in AI yet there's a shortage of people going into manufacturing where these techniques are being pioneered today, or at least the potential is enormous. When is that going to change? When is this lag going to catch up? When are we actually going to realize what a fascinating industry? [00:35:00]
[00:35:00]I guess you're in several industries, but just let's take Manufacturing, fascinating set of challenges. They can't all be solved by AI, but it certainly in your thesis, they can really improve and it's exciting.
[00:35:14] Michael Zolotov: [00:35:14] So I, I think that things are happening. I think they're happening maybe slow, slower than we would have wanted, but eventually these are traditional markets.
[00:35:23] And if you take, we have a client that ships every day, a million tons of iron ore to the entire world. So when you think these gigantic operations, it's not something that you can change over a day, but when you actually go and look through the years, you do things that you do see that things are changing in factories.
[00:35:42] The OEO factories that they do increase because of AI. I really do. It is changing slower than we would have wanted, but it's because, it's so massive and it takes time to change things. And here in Israel, we're the startup nation. Our goal is to try [00:36:00] and solve everything for technology.
[00:36:01] That's why we have it so many people here are going into the directions of deep learning and AI in general. And we're trying to maximize the impact and, really push our clients and help them as much as possible many times is very clearly showing them how this algorithmic superiority translates into dollars into better throughput into lower emissions.
[00:36:25]It's very clear and this process take time. We're trying to accelerate it as much as possible.
[00:36:33] Trond Arne Undheim, host: [00:36:33] Yeah, Michael, I was just recently in Tel Aviv. I guess recently it was before the pandemic but anyway I worry for you. You seem very busy. You have to go to the beach on the weekends.
[00:36:48] Michael Zolotov: [00:36:48] Good to hear. Look, I always have to look out for founders and make sure that you get your quality rest. But it's been great to speak with you and thanks for enlightening us. And I look forward to welcoming you for the [00:37:00] panel. Next week we'll be taking a peak at the startup nation and all the exciting startups that are going into it and going into manufacturing.
[00:37:08] So thanks a lot, Michael.
[00:37:10] Thank you very much Trond. It was a pleasure being here.
[00:37:13] Trond Arne Undheim, host: [00:37:13] You have listened to episode 19 of the Augmented podcast with hosts Trond Arne Undheim the topic was machine learning and manufacturing, and our guest was Michael Zolotov CTO and co-founder at razor labs. In this conversation we talked about where we are with machine learning and AI for Manufacturing at what are the main techniques what's possible now and what will be possible soon? My takeaway is that machine learning is definitely entering manufacturing over the next few years. Already. Interesting experiments are underway to do simpler things, such as prevent future downtime using sensor data already being captured by advanced machinery.
[00:37:57] Pure machine optimization can [00:38:00] only get us so far though. The real potential lies in complex business process optimization and simplification with Augmented frontline operations technology plays a part, but clever workers, operators, and engineers will have to make intelligent use of the technologies available they just blindly implement for that. We need reskilling, always learn. Thanks for listening.If you like the show, subscribe at Augmented podcast.co. Or in your preferred podcast player and rate us with five stars. If you liked this episode, you might also like episode 27 industry 4.0 tools, episode 13 gets manufacturing, superpowers, or episode 14, bottom up and deep digitization of operations.
[00:38:54] Argumented upskilling the world course for industry 4.0 frontline [00:39:00] operations.
CTO & co-founder at Razor Labs
Co-founder and CTO at Razor Labs, Co-founder and board member at Axon Vision, Axon Pulse, and Compie Technologies.
Creating innovative AI solutions for the most complex challenges and directing their business development allowed Michael to serially establish various prospering enterprises.
Michael advises to multi national corporations and is considered a thought leader in the fields of AI and Deep Learning.
Michael was a participant in the prestigious Talpiot program and a Captain in an elite R&D unit in the Israeli Defense Forces.
Michael holds a master’s degree from Tel-Aviv University, having researched lung cancer detection, using Deep Learning methods.