Machine Teaching for Better Machine Learning with Mark Hammond - #43

Key Takeaways

[Music] hello and welcome to another episode of to animal talk the podcast where I interview interesting people doing interesting things in machine learning and artificial intelligence I'm your host Sam Cherrington it's been another exciting week here at 1200 headquarters just a few days after hitting the 500,000 listens mark thanks to you once again we learned that at least a few of those listens came from a certain Mark Cuban and yes I mean that Mark Cuban speaking at a conference in New York City mark mentioned that he turns to this very podcast to learn about and keep up-to-date on advances in artificial intelligence mark if you're listening we love you man thanks for the shoutout the CNBC article that covered Mark's talk and mention this podcast focused on his fear of AI and what it might bring in the future as you might imagine this is a topic I've got some opinions on and I respond to the article and Mark spheres in this week's newsletter if you're not already receiving it heads at will Malaya comms last newsletter to sign up and I'll make sure you get the current issue last week we announced the first of two winners for our artificial intelligence conference ticket giveaway winners receive the bronze pass to the conference which grants access to all keynotes and sessions our second winner is Richard s from Brooklyn New York thanks again to everyone who entered the contest if you didn't win this go-round but would like to join us at the conference use the discount code PC twimble for 20% off of registration will link to the conference in the show notes which you can find at twill Malaya com slash talk slash 43 the first twin will online meetup was last week and was wonderful the focus of the meetup was the CVP our best paper award winner learning from simulated and unsupervised images through adversarial training by researchers from Apple the idea behind this paper is this consider a problem like I gaze detection you've got a picture from for example a cell phone camera and you want to determine which way the user is looking generating labeled I gaze training data is hard and expensive generating simulated I gaze training data sets is much easier and cheaper though and can be done for example by using something like a video game engine the problem is that the simulated I gaze images don't look close enough to real images to train a model to work effectively on real data this paper proposes using a generative adversarial Network to train a refiner model that can make simulated I gaze images look like real I gaze images while preserving the gaze direction thanks again to community members Josh Mennella who did a great job presenting this paper and to Kevin Mader for walking us through a tensorflow implementation of the model you guys are just awesome we're working on getting the recording posted for those who weren't able to join us live if you're signed up for the Meetup or the newsletter you'll be notified when it's available if you'd like to join the Meetup head over to twill Malaya comm slash meetup to register next month's meetup will be held on Wednesday September 13th at 11:00 a.m. Pacific time and we'll post the details of the program shortly before we get to the show I'd like to give a shout out to our friends at wise io @ GE digital for their sponsorship of this industrial AI podcast series hopefully you caught last week's show featuring Josh Blum vice-president of data and analytics at GE digital we had a great discussion about how to incorporate physics-based information into machine learning models among other things for more information you'll find that shell at Twilio comm slash talks last 42 and for more information on wise io @ GE digital visit wise io and now for today's show if you've listened to any of the shows in the industrial AI series you've undoubtedly heard me mention our friends over at Banzai I'm super grateful to Banzai for taking the lead and sponsoring both the estriol a I podcast series as well as my paper on that topic well today's show which concludes this first season of the industrial AI series features my interview with bonsai co-founder and CEO mark Hammond our conversation centers on the role of what he calls machine teaching in delivering practical machine learning solutions particularly for enterprise or industrial a I use cases I really enjoyed this conversation with mark and I know you will too and now on to the show all right everyone I am here at the offices of bonsai with Mark Hammond the co-founder and CEO of the company mark welcome to this week in machine learning in AI thank you for having me happy to be here yeah I'm super excited to have you on the show folks who are regular listeners will know the name bonsais without a doubt because you guys have very graciously sponsored my research into industrial AI and the podcast series and I'm really looking forward to digging into with you what you know what industrial AI means for bonsai but before we dive into that why don't we why don't you spend a few minutes telling us a little bit about your background sure so my background is actually originally very technical I started programming very young and ended up working at Microsoft's while I was still in high school on Windows itself Windows 95 and the first mini versions of Internet Explorer so definitely hands-on coding on the products themselves oh wow my passion though has always been artificial intelligence so even while I was there I knew that was what I wanted to do and I decided to pursue a course of studies in computation and neural systems at Caltech so I was working at Caltech I'm start working at Microsoft attending Caltech and it was great in all regards other than that it happened in the late 90s which was a fantastic time to work at Microsoft and like not the best time in the world to be in the field of AI like I went to part of one of the AI winters behind so I found myself when when I was completing that those studies kind of faced with well how do we how do we use all this stuff in the real world and at that point in time it was really well do you pursue a course in academia right is that you go to academic route or what else do you do because it's harder they I went to and so I kind of decided at that point that because I have this strong emphasis towards applying this technology in real-world scenarios if I wanted to do that I was going to need to get some of the not purely technical skills and so I so I decided okay I got to go try product management develop our sales and marketing etc etc and so I pursued that course I did find myself at one point back at Microsoft this time in sales and marketing I was one of the developer evangelist who was was out pitching net1 net1 brand new and getting everyone on the C sharp bandwagon so that was okay that was a lot of fun and fast-forward to today and the market is now in a great place where the technology is there capable it's the right time to start looking at applying these technologies to our real world industrial and commercial enterprises and looking at those use cases and I had come to the insight which led to bonsai and it was born through having gone through the academic track having gone through the pure business track having looked at trying to apply in lots of different contexts and ultimately coming on one very simple realization it's one of the one of these things where you you look back and you say well that seems obvious in hindsight but until you think about it it's not apparent and that's that no matter how good we make these algorithms they could be as good or better than humans at learning we will always have to teach them you have to teach them something it's a learning algorithm it's kind of by design asked a lot and there wasn't a huge focus on how you actually teach something we spend so much time in this field focused on on the machine learning algorithms themselves that teaching is often an afterthought that was the spark that said look if we're going to be able to solve these real world industrial AI applications the subject matter expertise the ability for people to define what they want to teach and how to teach it that is an area that is ripe for enabling the technology and that's what led to founding bones I am sitting here today awesome awesome it's interesting that you came across that realization I I tend to find the same thing that the emphasis is on kind of the machine learning and you maybe put another way the way people tend to think about this teaching process is throwing a bunch of data at an algorithm I guess it's kind of analogous to like throwing a bunch of books at a kid and expecting them to learn on their own yeah yeah absolutely absolutely the one the one that I often uses was with my son he's learning how to play baseball and and I tell people I don't take him out to the backyard he's five right so I tell people he's five I don't take him out to the backyard and throw fast balls at him because that would just be that's just cruel why would you do that and yet no one even pauses for a moment when we're like well we're just going to throw giant datasets at these machine learning algorithms and I bet if I threw a million fast balls at my son he'd figure it out eventually too but it's just not a very efficient or effective mechanism for teaching so yeah and probably be pretty painful for your son true-true not not very responsible for me as a parent yeah so maybe we can I guess we let's just dig into this so what when you talk about you know machine training what does that mean or machine teaching what is it what does that mean to you so machine teaching really boils down to actually looking at the art and science about how we teach things distilling out the abstractions that are there and then providing the appropriate platform tooling etc that developers we've come to expect in order to be able to properly construct the solution to a problem in that context so that's at a very high level but at a concrete level what it means for us is we need to give you a very formal way because it's still a computer program at the end of the day we need to give you a very formal way to specify what it is you're actually trying to teach and how you can go about phasing that teaching in a way that follows again using examples we're just talking about you don't want to have it be just throughout massive tros of data at the system how you teach it can be broken up in ways to facilitate acquisition and mastery of these concepts you're trying to teach so really at the end of the day it's about empower developers to work as teachers and giving them the ability to do that in a very formal structured way you can think about it if we're doing this in very humanistic terms you talked about textbooks right so a textbook has a curriculum that's set out in it here is the table of contents all the concepts we want you to master we're going to go through it in this order here's spaced problem sets that ramp and difficulty level you go through all of these different ways of structuring the textbook in order to try to help guide students as they learn things we do the same thing we just do it in a very formal structured way we give you an actual programming language where there's no ambiguity about you know what you're trying to do it's less freeform textbook and more much more program that you're creating maybe you can make that more clear by walking us through an example or you know pick a use case and maybe talk about how you would apply this to the use case yeah absolutely so if we look at a robotic system as an example so you have an industrial robot perhaps it's a robotic armature being used in a manufacturing setting or a warehouse setting and you want that armature to be able to create conduct various pick-and-place operations maybe it's doing a palletizing operation something of that nature now when you want to have the piece of equipment learn to accomplish this task you can break it up into the constituent concepts that matter it's not about I want you to perform the entire task let me demonstrate it for you it's about okay you need to understand the concepts of moving between points here is where you're currently at in space here is the target you need to get to how would you drive the motors to get to that that target that would be an example of a concept now for concept like that if you're looking at a commercially available industrial robot there are very good controllers for that kind of thing already right they've got the inverse kinematics all worked out and they know how to do it not just efficiently and effectively but in a way that's going to maximize the lifetime of that equipment you don't want to unnecessarily drive it too quickly so you're going to cause it to fail faster than you want real-world concerns when you're using these kinds of robots but then when you get to a grasping concept so now we need to actually grasp the item that we're going to be picking as part this action that is much more complex right and you have to deal with rigid bodies and soft bodies and different packaging materials and all these different kinds of aspects that come to play teaching the nuances that go along with that that's where you can start to take your subject matter expertise and really bring it to bear there are people within all of our organizations who already know a lot about these facets of things and they're not programmers per se they're they might be a mechanical engineer and they might be a chemical engineer depends what kind of problem you're tackling but they know a lot about that area and so if you ask them if I asked you to teach a human to do this what would you do what would you actually what were the what are the concepts you want them to learn how would you show them those things what how would you test whether they got it right or not you they usually can tell you actually because that's their field they know that pretty well and so we just give them a mechanism to capture that so in the context of grasping for example since we're just walking through one of these use cases you might say well actually if I'm teaching a human to grasp something I got to rewind a lot until back when they're toddlers right but what do you do you use large-scale gross motor skill objects you get them really close to their hands you have set their hands on it mostly and then you let them go through the motions and you it's hard to think back in many of these cases for these very simplistic motions because to us it's simple but if you watch children doing it they have to learn to but you'd break it down and you literally break it down into those kinds of things where I'm going to teach you gross motor skills I'm going to teach you in these simulation environments so that you can experiment frequently and you learn these concepts related to grasping so you learn the concepts related to grasping you're leveraging the pre-existing concepts related to moving between points you want to teach concepts related to stacking or placing orientation valid grasps so that you can orient parts in appropriate ways for fastening them etc etc all these kinds of things that happen when you're doing real-world industrial robotics you can break the problem down you break it down into these constituent concepts you design a plan for how you want to teach it typically that will involve some simulation environment in conjunction with some real-world physical environment and then you define what that curricula looks like to teach it and then because you've done it in that way the system can proceed to try all the various areas that it can explore to teach how that works and mathematically if we're looking at the low levels on the math all that's really happening is you're constraining the state space the system needs to explore that's what in practice that's what's actually happening but it's happening in this more naturally expressed way that a subject matter expert can readily latch on to and and work well with so that's that's an example in in a robotics context if you look at cogs or they'll not go too far because there's so much yes you're had an example to unpack sure I think the first thing that jumped out at me was you kind of described these two different types of concepts one that you know a lot about and you can help me refine this language and another that you know you need to teach more abstractly so for example you know in moving you know in kind of the the macro movement of the robot arm from point A to point B you know it's a well understood problem you've got the inverse kinematics you mentioned yeah I get the impression that you know we've talked a lot in this series about kind of reinforcement learning from a research and academic perspective and one of the you know the problems are I think in that domain not decomposed in this way and so I think what I heard you say was you'd be kind of crazy to like throw a bunch of data and like have the robot try to figure out on its own the best way to move from point A to point B when hey we've already done that yeah here years and years and we spent a bunch of time perfecting the way to do that so you know part of what I hear you talking about is kind of is an idea of modularity yes I'd written in these approaches that I'm trying to like get a whole lot of stuff at once sure no problem one thing I wanted to dig into is like you know compare/contrast you know what you're doing with you know the way some of the things we talked about on the podcast kind of academic approaches to reinforcement learning and so one is this idea of modularity another is you may be kind of elaborating on this this idea of constraining the state space in a way that is easily expressed by humans like I think constraining the space the state space is a huge part of you know this process even from an academic perspective but my impression not being an academic and at the start of this field but my approach is or my sense is that their approach is constrained in the state space mathematically right as opposed to conceptually is that fair that's fair at the end of the day it all does boil down to a mathematical constraint it's just how you enable people to express that and by virtue of building a system where you're expressing it in this more natural way for a subject matter expert who's actually working on these problems you can get at the underlying math by allowing people to express it in these more natural terms perhaps an analogy here would be would be useful if we look back at old programming systems right so in the in the late 90s when I was at Microsoft Visual Basic for building desktop apps was everywhere people used that all over the place and it was very popular because it enabled people who had subject matter expertise you know I'm running my veterinary clinic I'm doing whatever it happens to be to build the applications they can eat it they cared about because they didn't worry about comm interrupt capabilities and all these you know low-level stuff they worried about can I build a form and can I put the right components onto the form that I care about and tie that back to a database in a way that doesn't require me to go become an expert in assembly language and low-level binary interface technologies this is the same kind of thing it's about building the right abstraction at the end of the day and so even if what our technology is going to do a lot of the bonsai platform itself is going to take all this code that you've provided and compile it and yes at the end of the day it's a big mathematical constraint it's not that fundamentally the technology is different somehow it's the same it's just that we're allowing people to express it at an appropriate level of abstraction where it's now framed in the context of the subject matter it's framed in the context of a business problem you're trying to solve and that's very powerful because it takes it so that your data scientists can still play the role that's appropriate for them to play your programmers can play the role that's appropriate for them to play and the subject matter experts can participate and actually teach it the intelligence that you actually want the system to exhibit typically what we find in a lot of these environments is if you have true deep expertise in machine learning and data science that is whole on field alright and that the people who have that and if you look at the intersection with the people who have the expertise in building manufacturing equipment or optimizing supply chain facets rare right is very rare that that comes that they overlap so we have to provide as an industry we have to provide a way to enable all of these disparate skill sets to work together and we have to focus on the skill sets that are already within these organizations or we're never going to solve the real-world problems and so that's that ultimately is where we're getting at with this technique yes it's about decomposing the problems and it's about decomposing the problems in a way that allows these subject matter experts who know about all the different facets of the use cases they care about to really come in and say I need to teach you about this so if I look at look at real-world examples we're doing a lot of work with Siemens at the moment as a they're one of our customers and if we look at their manufacturing equipment that they come to us and they want to talk about adding intelligence to I'm an expert at platforms and artificial intelligence in all this time not an expert at CNC milling all right that's not that's not that's not my area but when they come to us and they say well here are the real-world problems we face when we have these gigantic pieces of manufacturing equipment and they can have an expert get on the line and they that expert one say well you're going to need to understand this facet of friction compensation so on and so on and you know areas that I know that nothing about frankly does that on a personal level but they can tell us all sorts of things about that and we can work with them to say alright well how do you go about this now how would we break that down into thing we can measure into a set of concepts the system can learn and it's not about how do we how do we craft a you know Pig controller to solve this problem which would be a traditional way to solve this problem in an enterprise context it's about how can we tell whether it was correct or not so this is where the reinforcement learning part comes in you don't have to be able to specify the controller at a mathematical level you have to be able to assess whether the behavior was what you wanted and the ability to break down that behavior into components so that you can assess for each of those constituent components whether or not that was what you actually wanted so elaborate on that do you I mean ultimately in these use cases you're trying to are you trying to create the control or are you saying the controller already exists and you're trying to identify the right parameters for the controller are you trying to say the controller and the parameters exist and you're just trying to do some kind of validation so all of the above actually we run it unfortunately the answer is that we see all of these scenarios so there are cases where what you have is that you have an existing controller and what you're looking to do is to identify deviations right so you're really trying to figure out when you've deviated and you have some condition that was not anticipated and you want to be able to deal with it appropriately so that's something you definitely run into there are areas you run into where you have existing controllers and you want to enhance the capabilities of those controllers beyond the already well-defined characteristics so you run into that as well and then there are areas where you people are still operating things by hand so it is not uncommon we can look at CNC machines right and we were just talking about them to have expert human operators Manning these machines because the value to the business that is using them and creating the part let's say you're making a large scale aircraft part that part might be a couple hundred thousand dollars right just for that one part and it might be a week-long operation to mill that part right you're not going to just turn it over to your G code script run okay and you're going to have someone there to make sure everything's going as you expect if you if there's a mistake it on day two you want to stop it on day two so that you're not wasting tons of time and money as you're going through that at the same time one of the things that I hear over and over again is that there you know from the perspective of trying to apply kind of modern saying you know business and engineering practices to some of these industrial environments a lot of what the subject matter expertise is is hey when I hear this machine kind of sounding an octave or two higher and Pitt's or something like that I know that you know we're probably going to lose a bit or we're going to probably damage the part or something like that there's a lot of a lot of art in addition to the science how do you how do you begin to capture all that so that's that's actually an excellent point so the beauty of the modern machine learning technology is its ability to detect nuances there where the human expert the subject matter expert can say yes I hear it and when it sounds off then I know this is about to happen and you can say well what are you listening for and they're not experts in acoustics they're not going to sit there and tell you well it's exactly this is the kind of sound right it's more like well I just know what I just know what I'm listening for I've heard it before and so the traditional mechanism would be go through label data set etc and you can you can still do that there are techniques you can use in simulation as well to model those environments but in practice the benefit you get from modern machine learning technology versus expert systems saying if we go back to the 80s yeah is this flexibility so if I can use another analogy which might be intuitive to to a lot of people if you play a sport and you really enjoy playing that sport and you practice and practice and practice and you get good at it and someone comes to you and they're like wow you're really good what is it that you're doing what am I missing so that I want to be good at this for - alright oftentimes there's an amateur you don't know you're just I just practiced a lot I got good at it and if you go to a professional coach or a professional athlete they can tease it apart they can well actually when you were doing this motion you'll note that you arch in this way and you know they can get into all the subtleties in the new ones that's why they're a coach role or an hour Pro and so humans as a learning system if we look at ourselves as learning systems we have this remarkable ability to be able to exhibit intelligent behavior regardless of our ability to explain all of it right and modern machine learning systems are like this so if you take a deep reinforcement learning network neural network kind of approach and you apply it to a problem and you say here's the correct behavior just look at it over and over and over again whether that's because it's getting acoustic data and it's listening and you're telling it whether or not the part was about to break or if you observe that the part broke and now it's learning what the subtleties and the acoustics are so that it can have that same sense that the expert operator did it can do that right that's one of the benefits of technology but the more you know about the problem itself it enables you to decompose it into those bits and so you're not forced you can always use the technology and get it to the point where again you're the amateur athlete and you you just learned because you practice so much or in the industrial case I have my system and it's actually monitoring the acoustics off of the equipment and it's learning to detect what it sounds like when a part is about to break it can do that that's fantastic but then if you have that subject matter expertise you can really decompose the problem you can get a lot of benefits because you can teach faster you can now have the predictions that are made explained so that the system can make more nuanced and more accurate behavioral decisions and really getting that subtlety and nuance allows us to build and capture more knowledge and build more sophisticated systems it's kind of like with expert systems you are totally rigid here are the rules I'm going to infer all this behavior it's been powerful in that sense but very constrained it was not very flexible and now the pendulum has swung the entire opposite way we're all the way at the other end and you have your machine learning systems and it's throw lots of data at it and it's going to learn to predict something and great it makes great predictions but no one knows why right so before totally explainable completely inflexible now totally flexible not explainable and by virtue of using a machine teaching approach like the one that we've outlined it's no longer black or white you get this nice continuous gray area if you don't have a lot of that you can provide the system can still learn and that's okay the more of it you provide the more explained ability you get the faster it can learn the more nuance you can add the decisions you're making and it just opens that up and that's so it really allows us to tackle these problems at whatever level of subject-matter expertise explain ability is appropriate for what you're trying to do okay so what I think what I just heard was you will you talked about explicitly the spectrum but when a company is using your tools and building you know building a solution based on it they've got the ability to you know you can start by at the highest level by throwing lots of data at the problem and not building you know building constraint into the system or you know conversely you know articulating the concepts you know breaking down the concepts that compose the system you know or you can do that to some varying degree of detail so it sounds like that's probably one of the kind of architectural design decisions of someone that's implementing this like how much decomposition do we you need to go to and is that what are the factors there is a primarily performance is it actually it tends to be very iterative so technically typically what ends up happening in a in a real life engagement is they will first start with the simplest possible model which is there's only one concept and I just having I'm doing the classic reinforcement learning thing here's the environment for you to go explore go explore it and that might be a robotics environment it might be supply chain simulation it could be any number of things but just just explore and see what you can figure out typically that will learn something mmm not as much as one would hope but but it's offering and then you say okay well let's see what would happen if I taught it about this and so you add some conceptual block into the system and you break it down into teaching that and it may or may not help it's not always a given that it helps oftentimes we'll get feedback from people who are more on the academic mindset of the whole point of deep learning was to get away from specifying these things right then right that was the whole point why are you doing this what happens if the person specifies this model and they're wrong right that one of the benefits of deep learning is that it's not reliant on our presupposed conception of what the model should be right how do you cope with that and for us it's like well that's beautiful actually because when you start to break it down and you decompose the problem and you do it in this iterative way and you see whether that supports faster learning whether it supports better explanations better reuse better generalization all these different factors you might want to optimize for care about you learn about your own model and so if your conceptual model is that the friction compensation is super important for this manufacturing process and you go through the motions and you receive it well the system is learning to make predictions and it's compensating just fine but all the things I taught it that I thought I knew aren't being used and the system can come back and tell you this it's well you taught me this concept and in fact I never use what I learned I'm always doing something else that's instructive that tells you hey this model I thought made sense there's something better because the system has learned the correct behavior and it's not using what and and this happens at all levels of granularity so in practice it's this iterative process people start at the very simplistic model they start to add more and more of the model that they believe is correct it's very rarely a here is the 120 concept model that I think maps to the problem I'm trying to solve and I'm going to go build the whole thing in one one go and go from there it's much more of a iterative refinements and expansion of the model so that you can have more nuanced covered and more subtlety covered and learn about your model in the process doing okay one of the things that you guys talk a lot about is the notion of explained ability yeah you just went into that very important for customers in this space for a variety of reasons I don't think I previously understood that it's this granularity of defining the concepts that really gets you the explained ability yeah so we elaborate on on that yeah absolutely so there are a lot of techniques that have been published about how you start to peer into the neural networks and try to tease out what is what is actually going on you got lime in a lime and all even there's tons there's tons of these things our approach is different as you said because where it's it's not magic in our system it's not like we're going to have some amazing new way to peer into the neural network and tell you what this group of neurons meant right right that's that's not what we do what we do is we say we're going to let you break it up in this way and let's it's always easiest to frame these things in real-world scenarios so let's say you are building us supply chain logistics routing system right anyway you have got your real world data so you have all the telemetry coming from that you have simulation models that you've built in some discrete event simulation like that so you've got all these all these different facets you can pull on and the model that you use takes into account and grow very coarse things like the weather and seasonality of goods and perishability of goods and not all these kinds of things that you might care about and you might have more fine-grained concepts that you're teaching it about the composition of your fleet you know and the land routes that are viable for you to follow and so I like all these kinds of things so you know you can teach all of this stuff what's important to emphasize here is that it is very rare for a company that we work with to come back and say and now we're turning it all over to this automated system we trained go that doesn't like media that's still there's not a level of comfort there yet what happens is there's still a human and the human analyst is sitting there and making the ultimate decision and they're using the system to provide decision support yeah and in that context if you built I don't care sophisticated is if you've built a very elaborate neural network or maybe you used some other machine learning or AI technique to build a system and it comes back and it says I think you should have truck 17 which is currently in Hoboken depart now go right so what I think you should do the analyst is going to sit there and look at it and say why okay that's nice that you tell me that and I understand that you have this visibility on massive parametric space that I'm not perhaps aware of as a human right but but I'm still not comfortable with the fact that I have no context into why you're telling me to do this that truck is only two-thirds full and you know the next truck that can hit that location is not available for two days okay I don't think that's a good call whereas if the decision support system comes back and says I think you should have this truck leave now and I think you should have it leave now because and then it frames that prediction in the context again not magic but in the context of all these concepts you've defined that was your own model so you as that analyst who is sitting there and presumably is either very familiar with that model or part of defining it we're going to look at that and have some rich context to say oh right so we taught it about these overland routes and it sees that there is a storm coming and those routes are going to be cut off and that's why it's telling me to do it now even though there's not another truck for two days whatever you know whatever it happens to be but it can frame it in that context that gives you much more power and as an analyst to make that decision in confidence in the results it's got an you get your audit trail this is why this is why this happened and so even if those models are are not ground truth right it's not this is not exactly the state of the world but it is the model that you use to run your business it is the model you use to drive your you know your robotic systems or whatever it happens to be you want to have that level of explained ability and that's what you currently use to frame it and so it's not we were taking a lot of the magic out of the eye but giving you the flexibility that you could have a note in there where the system comes back and it says I think you should have the truck depart now and everything in the model I'm not using any of that if you just use that model the model would not say make the truck depart now but I've also learned by virtue of looking at all the real world telemetry and it that I believe making that prediction now is a good good prediction then the analyst is going to look at it and say well I'm not really comfortable doing that but let's look at what happens let's see if that was in fact a good prediction that comes out and that tells me I need to go back and enhance my model my model is now deficient in some way and then you can iterate and you can keep working on it so I'm specifically deficient in its expression of these concepts yeah right it has a blind spot it's actually blind spot isn't really the right way to say it because the model doesn't have the blind spot it's the it's the lack of decomposition you can't explain it it's it's like going to the human expert and saying why and having them say I don't know it's just this is the way I do it and it works right so yeah humans have this interesting I'm a student of human nature as well and humans have this interesting facet where we conflate the ability to explain something with the ability to justify something and so it's important as you look at these systems and say well are you actually explaining my interview Jose or make this decision and that's really what happened or are you just looking at what happened and now you're justifying it ah a lot of times it's the latter it's not the former that's human nature and that's just that's just the way we are but when it comes to industrial AI and really applying this technology you really want it to be the former and so in certain circumstances you want to go back later and leverage the human strength which is to say system predicted I should do something that was out of the bounds of my model there's a gap assuming that it was the correct behavior then there's a gap how do I try to fill that gap and then your ability to justify and come up with creative explanations for what that might be the ability to have your data scientists dive in and really tease apart what happened and try to refine that model becomes very powerful so it is a very fluid process it is not a right once run forever proposition it is a right many times continually refine and learn as you go along and get more getting greater and greater and greater ability to explain what is actually happening as you go through that process that's that's the nature of the beast one of the things I found in my research and articulated in the industrial AI paper was an emerging maturity model in the way people are looking at deploying AI and in the enterprises generally I think but particularly in these industrial types of situations where it's just as you described right there's this fundamental issue of trust and that Trust is I think multifaceted part of it is you know repeatability part of it is explained ability there are a bunch of other factors and as people are gaining this trust the first thing that they want to do is you know point these systems at some process and some you know system and just have it help them monitor it and tell them new things about it and kind of surface new insights and then as they gain some some trust that hey it's actually providing me interesting information you know maybe it should tell me what to do right then they kind of flip the switch or you know allow it to optimize and you know they actually you know becomes a decision support system and way you described it and then there's this further you know stage which is actually you know the next switch which is just do it you know make it so right so control and I'm wondering if you see that same progression among your customers and you know what are some of the other factors you know that compose what are some of the most important factors that compose trust beyond explained ability so we definitely are starting to see some of that but I would emphasize be starting to part and you yourself described there's an emerging capability maturity model that people are using as a factor here and I would wholeheartedly agree with that but the state of the market Knight right now really is it's like its 1995 all over again and I think there was this cool visual basic thing well it's right up in its face it's like there's this internet thing and everyone's like wow I really need to do something about this internet thing but a lot of people the maturity of people's ability to do that is all over the map right people don't even know where to start they know it's important and you have true deep experts who are doing it and so we see all of that and as you engage with customers and they're looking at the maturity of their trust in these systems and trust in their own models and own systems they've built so that they feel more comfortable turning it over we definitely see that you can look at as an example of Tana miss v achill so we talked to of course lots of people about a ton of C equals dot area and there was a period of time not long ago where no trust whatsoever in having the systems make control decisions using the technology for perception and identifying whether there's a bicycle on the road or a cow or something like that okay we're comfortable at that level right now not steering the vehicle like no that's not not yet and you see some of those organizations now getting more and more comfortable with him and but really it boils down to how well baked is that the system that you've built they can be at the point where it is just you still want a human in the loop it will get to the point where you want a human to validate what's coming out and then yes you will get to the point where frankly you will have the system make decisions in automated way and you do that because there's now sufficient trust and confidence in the system that it's going to do the right thing that doing the wrong thing is now an exceptional activity and they'll still make some mistakes but it's rare and when it does make a mistake you can capture that and you can use that to further refine the system so yeah I think that's a natural progression we are starting to see that but frankly at this point in time the market is really all over the map and so tons of experts in a room we will run into that with a you know with some frequency and you get into the occasion where people are just dipping their toes in the water and everywhere in between so it kind of depends on the particular zuv the customer and how forward-looking they are and how much resource they have to allocate towards exploring various facets of what they can do but buying further a large part right now there's a lot of desire to have explain ability to have that audit trail if you will so that people can go back and test things the more you get to automation towards the automation end of the spectrum and more people want to be able to look not purely at the audit trail but in generalization okay so if you look at control systems for robotics so which we're talking about at the beginning am i teaching the right concepts such that it will generalize the behavior in many scenarios not just the one I'm teaching it about right so there tend to be more towards that end of the spectrum of trust and automation so that's how they could perform their tests it's I didn't just learn how to grasp in a way that is proprietary to do it did I learn grasping in a generalized way such that if I present a variety of objects in a variety of scenarios it will still do the right thing and consequently you see you see everything you see you see everything on the spectrum at this point in time just to follow up on that last comment you made about the generalization I imagine that's got to be driven by a business driver you don't want to have an overly generalized system because that's going to be more expensive than what you need to solve your problem a hundred percent I agree yes do you find I don't know what the questions employees I know it's very totally you're totally accurate though okay if you have let's just keep it in the same vein it's easy to continue yet with that example so robotics okay so I have my robotic system what I have it doing it's been retooled we are part of a chain of its manufacturing chain we've just recently retooled for this chain I really care about this one operation I need to attach these two parts of whatever we're building that means that the grasp has to be in a certain orientation because it's only viable to connect the two things if it's I'm not covering up the part that needs to be right then you're going to be combined and so on all these kinds of things and there that's what they care about they're really focused on that because that's what they're manufacturing right now and that's what's economically effective and efficient and capable and if you look instead for the person who manufactures that piece of robotics equipment and so I have I'm a BB or I'm I'm a you know you pick your robotic manufacturer of choice I make robotic armatures I care a lot about the generalization because for me having it work with my god my my customers on a variety of context helps them go faster right and so that so that's kind of the breakdown we tend to seen out and then I think I imagine it also goes back to this idea of customer and market maturity right so you know my first few projects are going to be like proofs of concept and things like that and I was trying to solve this thing quickly and see if this is a viable technology but at some point you know project 3 or 4 or something like that I want to I really want to understand that you know if I'm going to invest in an approach or a platform that it can solve a broad swath of problems because there's real cost to total that investment totally and that's that's exactly what we see so it is a typical engagement for us at the moment would be to have a we have an early access program that we're actively working right now and customers coming into that almost always want to run a proof of concept as you said as a first stab we want to try to make it as aligned with the ultimate production use cases as appropriate but also appropriate duration and so on so that it's a spend of resource and people's time and money is aligned with pinning down exactly what you said is this technology something that makes sense for what we're trying to do and as people get that level of comfort then yes then you get more into the okay now let's expand the scope and we're doing a much broader use of this technology in the production environment and so on that is a progression that we see time and time again and I think that's that's true that's not just true about us I think that's true about the this technology in these industrial contexts in general mm-hmm we talked a little bit about about reinforcement learning and you know in some of the examples you gave you start high level problem you decompose it into concepts you know you know if the concept is like a leaf node you know somewhere under there you're doing reinforcement learning to kind of figure out that leaf node if you don't already have a well understood model like inverse kinematics yeah help us understand the relationship between what you guys are doing in the underlying reinforcement learning stuff like are you how are you how you architect in the your the networks how are you you know training them are there you know any particularly interesting things that you're doing to you know ensure that they're quickly trainable though is there any academic research that you have based your approaches on like what are the things that you think about short at interphase you're sure that's it that's a there's a lot of depths that we can get into there so let me start at a high level so we can frame everything and then well I'm happy to push down there at a high level the four bonsais platform in particular the best way to think about it is in relationship to a database so when you're building industrial or commercial enterprise application X for your company whatever it happens to be and it's going to work with data you're going to use a database almost certainly that's a very common thing to do and what is the database providing for you it's providing that level of abstraction so that you are not focusing on how this data is split across discs in the cluster and when you rebalance tree structures for searching and all this kind of like that all that stuff the database deals with for you and it gives you this nice abstraction so you specify the structure of your data and the kinds of questions you want to be able to ask of it and it can take care of the rest our platform is very analogous to that very similar now of course our abstraction is around this machine teaching stuff and so on but in principle when you're working with these simulation environments when you're working with real-world equipment telemetry you are nonetheless interfacing with all of the systems and you have to manage them so if you were building a system and you're going to go through the actual training motions and let's say it's a supply chain logistics system and you have a discrete event simulation model of that system and you want to Train primarily on that before you bridge to the real world telemetry data just so you get quicker learning and more repeatable learning then you have to manage those simulations and those simulations depending on what you're doing those can run very quickly in some cases if you're doing a computational fluid dynamic simulation that can run for hours or days and so managing all of that becomes something that matters and so in the same way that you don't worry about data spanning different disks in your cluster on your database we don't want you to worry about am i running ten copies of my simulation environment and where am i running them how do i reconfigure them between runs so that I'm maximizing the efficiency of the learning all of that is the kind of stuff that our platform manages for you and so in that sense there's a lot of low-level plumbing infrastructure stuff that is really valuable because you don't have to worry about it and it manages that for you but then when you dig down okay so that's the high level now we drill down a level great I've provided this but that's interesting stuff because at some point you know someone's got to set up you know models and actually train them getting to you know have that stuff all automated you know out of the box is a lot of work that someone doesn't have to figure out how to do yeah and in fact that's kind of the state of the art for a lot of this work if you look at the academic literature and you look at deep reinforcement learning algorithms in particular you'll find the apg networks and TRP o networks and questions around whether you should have learner memories or not and and is it a pond policy or off policy method because if it depending on which way it is you might have to throw out historical data you've cashed given what you're learning next and and this level of detail is great if you are focused on the mechanics of the learning and from our perspective it's the kind of thing you shouldn't have to worry about if you want to focus on what you're teaching and how the system should be intelligent at the end of the day what does the subject matter expertise is I would I would rather abstract all that and manage it for you and so by you know abstracting and managing it for you you know when you kind of punch into the details can mean a bunch of things that can mean you know actually we know about all this stuff we know how to figure out which is the best thing you know and we do that or it mean can mean what we really only do this one thing you don't need to have one choice and thus we've managed it for sure I sure yes so so it's more the former than the latter of course if you come back as you'd expect you never know yeah no it's a totally fair question so what does that mean in practice and let's let's use analogies again because it makes it easier to recognize let's say you're teaching the system something silly I wanted to learn how to play tic-tac-toe I pick tic-tac-toe in particular because everyone's played it and everyone understands if you're past a certain age that it's a pointless game because you'll never win miteta right the state space is not huge for tic-tac-toe there's a funny xkcd comic actually where where he literally maps out the entire state space in the comic like it here is every position you can possibly enter and what the correct move is once you're in that like you can never write and you just do it it's small enough of the state space you can do that we'll link to it in the show notes yeah awesome perfect I'll give you that link that's easy and the thing that's notable there is if you were teaching that and you did it in our system part of what the system does when it's compiling and it's generating the appropriate underlying networks on your behalf is it will just run the simulation for a while and it will collect statistics on what it sees and if it's running tic-tac-toe after it runs it a little bit and this is not take a lot of wall clock time it's just got to run a lot of iterations of the in this case of the game but it runs a lot of iterations of it it will see very quickly the state space isn't large and consequently maybe using a cue table is a perfectly reasonable algorithmic approach to solving this problem cue table being cute table being a specific approach for building a reinforcement learning Network where it's appropriate if you don't have a lot of options to the state space you cannot accumulate the entire if you can enumerate enough of it okay just a table is a perfectly reasonable thing to use and I mean that's a oversimplification of a cue table but this is a general so that's a general gist yes and fine why not use that that's an efficient system to use in this context whereas if I give it chess and you say go run some sample iterations of chess and just see what's going on it will very quickly learn that the state space is gigantic it's huge and using a cue table is not an appropriate choice in that context you need to use a different type of network and in fact depending on how deep that network is and how well you've decomposed it we might have to have some pretty deep layers and layer stacks in that network to be appropriate so that's one level of like how it decides which algorithmic approaches to use it also looks at the kind of data that are flowing through so if you're looking at chess you can hand me the data as a array if I'm looking at a robotic armature you might hand me sensory data which is you know just a collection of floating point values for motor torques and sensor detection and so on and if I'm looking at autonomous vehicle you might be handing me visual camera data right on I could be you could be handing the system any of these things the appropriate network to utilize is completely dependent on that data so if you hand me a visual data I should be probably constructing an appropriately sized convolutional Network right exactly it has spatial locality if you hand me audio data it has spatial locality in a sense but it also has temporal locality so I should be using different kinds of networks for that and all these kinds of heuristics and ways of looking at the environment and exploring and looking at how you deconstructed the problem they're all taken into account and so when the compiler outputs at the end of the day okay this is the network topology we're going to use this is the algorithmic structure we're going to use etc etc that's what's informing at all and on top of that we we ourselves are a machine learning system so it learns by virtue of having explored how to solve similar kinds of problems we've seen before how to tune hyper parameters and all these other kinds of things again levels of detail that the compiler should deal with you for you the platform should deal with for you I would I want to if if we've done our job properly it's much again going back to the database analogy it's just like that you don't have to tune any of those things you don't even have to know what those things are or what they mean you just have to be able to tell me what you want to teach and how to teach it and we'll build something for you is it the most efficient thing that you could possibly as hand-tuned if you were an expert and you how to do it yourself probably not and that's why would you still have database administrators and these experts who can go in and tweak all the I'm going to look at the slow log and go modify the queries and I'm going to all the things you do with databases same thing for us if you want to go in and you want to provide a node in the system where you tell us this is the structure this is literally the topology it should be for for this component okay we're not going to stop you that that's fine but if we give you the flexibility to decide where at what level you we want to do that that's better because it now lets you focus on the right people on the right levels your subject matter expertise can be focused on what's teaching how to teach it your machine learning experts and data scientists can be focused at that level and teasing apart when the networks are making predictions they weren't or it didn't fit in your model and you need to now refine it great that's where I want them to play your programmers have their role to play I need to integrate with all these simulation environments and I need to have the telemetry data being tied into the system and so on every participant in this process has a role to play and our job as a platform company is to make sure that they all have the right to lling that's appropriate for what they're trying to do and they're their part in the process and that we can have the discussion at the appropriate level is it at the level of the use case is it at the level of breaking that down is it at the level of the subject matter expertise or are we do we need to get down to the level of convolutional neural networks oh yeah network architecture is there a spec sheet or is there an enumerated list of like these are the networks that we support and how quickly this does that evolve are there limitations or so it's continuously evolving as you can imagine as a platform company we're always adding more we do have documentation up on our website so if you hit our main website and so that's all there yeah just link to the docs section there's documentation down to the protocol level so I have my own bespoke custom simulator which we do run into with some frequency and I need to be able to tie the simulator to the platform how do I do that right so there's documentation on that and so it just kind of depends on what level you want to get to but yes there is documentation all up online publicly available if you get asked by customers like how many different Network topologies or architectures do you support and yet that happens sometimes it does it depends on the times on the sub tank there she are ringing how depth that yeah they've gotten to so yes we've had people ask that and so we can talk about T RPO networks and DDP G networks and DQ n networks and all the different we know now we don't have to enumerate them all right now but yes you can you can start to enumerate all the different ones that the platform has baked in and have that conversation you'll also run into people who have been rolling their own so a lot of times they will get asked the question beyond which networks do you support I was just going to this guy said yeah we'll get asked okay we have been what will you will get asked the question well who else in the space should we be looking at and really there's two answers when it comes to deep reinforcement learning and applying that to solve these real world problems the first is rolling your own right that is actually we run into that very commonly and the second is very very vertically specific companies focused areas right so there are players in the space who only focus on deep reinforcement learning for supply chain only deep reinforcement learning for robotics at cetera right so those are the kinds of things you run into in the case of the latter if the solution they have fits your bill then that's that's great right they have a very specific focus we are a platform company so we're looking to allow you more flexibility in custom modeling and how you expand all that out so that it's just kind of a judgment call internally for over where you fall on what you want to do in the former where you're rolling your own which we run into all the time that's actually perfectly fine situation for us because the conversation quickly becomes well how much time have you invested in that all right and have you run into this set of problems so the over-under so you'll run into customers who they've decided reinforcement learning of the correct path to solve their problem they've been going down that path and building out the solutions they're spending a ton of time managing their simulations their timing is better and all these different assets assets of what we had just been talking about right and we say well you know we can we're not going to magically make the problem that you're facing about testing a good reward function go away right that's part of that's still part of the development process you're going to have to do that but we can take all the pain of managing those simulations off of your shoulders right like the platform can help you do that and so you run into those companies that are a little bit further down the roll-your-own and they're like oh wow I don't have to do this anymore that's that sounds great thought let me let's let's talk about doing it and then you run into ones that are just dipping their toes in the water they kind of say well I could just build you're built on top of tensorflow why wouldn't I just build it using tensor flow and we can start to enumerate all the reasons why but but generally we can say and these are the problems you will run into and some percentage of them will say I don't want to I don't want to deal with that let's talk and some percentage of them will say I got to experience that much I got experience and that's fine with us because we know that you'll you'll battle the basketball at Meade and it was a fastball as motivation would totally yes exactly you got to learn how to find the balance of where where it makes sense to roll your own and where and where it doesn't in actually I was going to I was going to ask the question more aligned with kind of that last way you expressed it and in particular you know tinterval is obviously a popular platform for this stuff like is it in either or no it doesn't have to be at all in fact we have a feature that we added the product recently called gears and for the developers in the audience this is an interrupt feature and for people who aren't developers per se it's what allows you to bridge this gap right it allows you to bring your current investments that you've builds whether it's a tensor flow or you built a perception model using open CV or whatever you happen to have done and add that to the system facet of building any of a modern machine learning system which is not just building the model but also deploying it in a production environment because there's actually completely separate problems are going to face in those two arenas and so you might have built these models but and want them to participate in the broader prediction pipeline right and you can do that very easily in our system using using the gears feature if we allow you to add that in so it's again going back to the database analogy you don't have to if you have that expertise and you want to do it or you've already baked a bunch of things and you want to leverage them we're not going to make you redo that work and we're not going to stop you from tinkering with the lower-level pieces if that's what you want to do it just becomes an option as any good platform should it's we're going to give you the ability to say I don't want to deal with any of that do it for me or I've done some of this and I want to add it or now I want to tinker with a low-level bits because I I feel like I've learned enough and I'm an expert and I want to do that now so you can you can play at any of these levels but it's not an either/or in fact if you bring a temps or flow model to the system it integrates really well our system is built on tensor flow a tensor flow based gear we can chain everything in all the appropriate ways if you have a Python function that you've written you want to call out to maybe you want to invoke a cloud API because you've made an investment in Watson or some other technology none of these are barriers and in fact this is part of the way the system was designed if we go back to the very very beginning and I talked about you have existing controllers that know how to do the inverse kinematics in it right all you need to do is move your robotic armature between point a and point B all right you should not be teaching that that is but and this goes to the academics right so you look at the academic papers they're going to talk about how you teach all these thing and that's great because it's talking to how we further and enhance the technology and I love that work that's great but if we're talking about practical real world application why you have tons of work spent on building it and you should just use it and that's a very simple simple example but if you look at autonomous vehicles the automobile companies have spent a lot of resource on building capabilities around assistive parallel parking and all these other things do we really want to go reinvent all of that it seems kind of silly we should just integrate that with the rest of everything else and that idea extends to your cloud-based API is your tensor flow model totally totally I thought that's right that's right it can't be a homogenous this is the only way well that's not that's not practical that doesn't work so this has been super interesting anything that you'd like to leave folks with well I would encourage the audience to take a look at our platform if they're interested in the early access program of course we would love to hear from you if you have a use case that you think is suitable we'd love to hear from you but just in general I would encourage everyone to start thinking about machine teaching whether it's with us or not with us the path forward for industrial AI in general is going to rely very heavily on the marriage of human expertise and machine intelligence you need both it's not enough to have one or the other you need both and so starting to explore in more depth how you're teaching the system don't just throw data at it library don't but that's that's level one right you need to move several levels beyond that so I would encourage everyone listening start to think about that start to think about strategies for doing that whether you're rolling your own or whether you're using a tool like ours that matters a whole lot and as practitioners using this technology a lot of our job is going to be teaching it's not going to be over time the tools will get better tensorflow will add more capabilities platforms like ours will become more prominent all of these things will happen just as a natural evolution of the industry and the part that will remain in all circumstances no matter what is how do I teach what I want the system to actually be intelligent about and that's going to be with us forever and that's very particular in idiosyncratic to our businesses and what we're trying to accomplish so really start to think about that that would be what I would encourage the audience to do great great and we'll make sure we point folks to your website and the EAP and some of the other stuff we talked about sounds great awesome thank you so much mark yeah thank you as well I really push you all right everyone that's our show for today thanks so much for listening and for your continued feedback and support for the notes for this episode including links to mark and the various resources mentioned on the show head on over to the show notes at Toma lay I comm slash top slash 43 please be sure to comment there with your feedback or question thanks again to our sponsors for this series Banzai and wise io @ GE digital I'm so so grateful for their support if you enjoyed this series it would mean a ton to me if you took a second to reach out to them on Twitter to thank them for their support at act wise IO and at Banzai AI don't forget to register for my newsletter at 1200 Icom slash newsletter and for next month online meet up at Twilio comm slash Meetup thanks again for listening and catch you next time [Music]