Understanding AI with Intuitive Visuals: Computer Vision for Handwritten Numerals
Tech experts without a background in AI get to see the fundamentals of the technology and learn about our start-up journey on the Innovators Show.
(We thank Mayah Schipper of Zetane for her helpful contributions to this article.)
Enough is enough with the weird robot imagery, especially of those brandishing weapons. Such common, sensationalist and often negative representations of AI are not only inaccurate; they also impede our abilities to educate decision-makers and the public about the important and fundamental components of this disruptive technology that is already impacting our daily lives. For society to reap the full benefits and avoid the risks inherent with innovation, it is in all of our best interests to gain a better understanding of what AI actually is, being sophisticated software capable of processing complex statistics.
Two of the co-founders of Zetane, Jon Magoon and Guillaume Hervé, had the opportunity to demonstrate concrete examples of what AI actually looks like as guests on The Innovators Live Show hosted by Montreal New Tech. This discussion series for technology entrepreneurs and startups from Montreal hosts local innovators that are pushing the boundaries of innovation to their extremes. Here we show non-experts the fundamental components of a computer vision model that can identify handwritten numerals; this is the underlying technology that makes it possible for us to cash a cheque using our mobile phones, for instance. The discussion also touches upon our entrepreneurial journey as a growing business in Canada’s booming technology sector.
See the discussion in this video or you can peruse a transcript of the conversation below.
Transcript of interview (with minor redactions)
Denis 0:15
What do you imagine when we say AI? Now I’ll try to imagine what AI itself actually looks like. If we look beyond the code and try to visualize AI working, what comes to mind, Eunice?
Eunice 0:27
A very technical image, like a bunch of code interacting with each other and interlinking different databases.
Denis 0:34
Yep, that sounds about the same for myself — to something very pie-in-the-sky, maybe a coder sitting behind his computer. I personally, you know, I don’t come from an AI background. So, to me, it’s a very much a mystery — “black box”. So I want to welcome all our guests, all our attendees, to Innovators powered by Montreal New Tech. It’s a show about people, and companies, organizations, who are changing the world through innovative ideas, technology, and science.
Eunice 1:02
We are your hosts: Eunice and Denis Luchyshyn. On the last episode of Innovators, we dove into the ethics and hype around artificial intelligence. Today, we want to build on that conversation and take a look at AI, but through a different lens. Today, what we want to do is to talk with two visionaries who set out how to make AI more accessible, democratic and impactful, especially for those who are not coders, but can actually see the value and potential of AI.
Denis 1:39
Awesome. Thank you so much, Eunice. So please, you’ll see there’s a little icon at the bottom of your screen. Click that to give a big round of applause to our amazing guests and welcome to screen Guillaume Hervé and Jon Magoon. So two of the three co-founders of Zetane Systems, developed in our own backyard here in Montreal. Welcome, Jon, Guillaume; how are you guys doing today?
Guillaume 2:04
Oh, very well. Thank you. Thanks for having us.
Jon 2:07
Yeah, thanks for having us. It’s a pleasure.
Denis 2:10
So what is the most exciting AI news you’ve heard this month?
Guillaume 2:16
The most exciting AI news this month, I think is the fact that like, on a monthly basis, we’re seeing so much investment going into AI companies. But what we’re seeing which is different than three, four years ago, is I think, is AI companies are more focused on what they’re bringing to businesses, and how they’re impacting, you know, the world around them. Whereas three, four years ago, it was a lot of just throwing money at different concepts and seeing what sticks. So I think there’s a more of a ‘focus’ in the world of AI today than there was, then it was just not that long ago, really.
Denis 2:51
That’s great news. It’s a very quickly developing industry. So it’s great to see that things are progressing. Anything on your end, Jon?
Jon 3:01
Yeah, just, I mean, there’s so much research that is always happening in the AI field and seeing some of the new models that are capable of achieving kind of state-of-the-art results, but use, you know, a variety of different architectures. So transformers, and then new models that also use a similar mechanism has Transformers called “attention”. That performed just as well as Google’s large models have been released in the past month with state-of-the-art image-recognition abilities. It’s very fascinating to see how the field is developing.
Denis 3:34
Very cool.
[Section redacted. Discussion was about setting up a game with the audience.]
Eunice 4:51
And while the audience joins us for this dynamic, I want to just give you a quick idea on who is on the show today because, Guillaume, I found out that not only you’re part of an organization that is doing so much within the Canadian AI community, but you’re also an author of a published book, would you like to tell us about it? Yeah,
Guillaume 5:17
Thanks for asking that. That’s interesting because it actually does lead into what we’re doing at Zetane. So I published a book on corporate entrepreneurship, which is “intrapreneurship”: entrepreneurship within companies. And it’s based on some of the work I had done; I had launched startups inside companies, one of the bigger ones became what is CAE Healthcare where we took the concept of pilot training, which uses a lot of modelling and simulation, and ported that know-how and technology to the world of healthcare. And so, so having done several startups within companies, I ended up not only launching that book, but that book led me to becoming a mentor at places like TechStars, and Founder Fuel and District 3 here just in Montreal, and a few others. So it was an exciting journey, which didn’t have a destination when I started it. But like all journeys, you know, they’re fascinating.
Eunice 6:12
And those are the more exciting ones. Very, that’s impressive. Thank you. Jon, just a quick question: you completed your education in Louisiana.
Jon 6:21
Yeah.
Eunice 6:22
Out of all the places you could have possibly ended up and how you ended up starting a company based in Montreal?
Jon 6:31
Yeah, well, you know, I’m, my background is in philosophy, not software development. So I’m a self-taught software engineer, and I’m a self taught AI and ML developer, and I’ve met Patrick and it’s kind of a long story. So I won’t go into it too much. But we all have a passion for making this kind of complex field more accessible to people. You know, my background is learning online learning, teaching myself through textbooks. So the idea that, you know, AI and machine learning, which is complex, but it’s definitely understandable, could be more accessible to a larger range of people was appealing to all of us. And I think we all share that kind of — that teaching background, and Patrick was a professor and you know, Guillaume already mentioned his mentorship and his writing. So it’s definitely something that we all connected on, and we were all interested in developing further.
[Section redacted. Discussion was about a game with the audience.]
Denis 11:20
Okay, so let’s dive into the real entry. Because there’s a lot of stuff I wanted to go over. Time’s going by fast already anyways. So let’s jump right in. So the reason why we got you guys here today is to talk about Zetane and the Zetane Engine. So you’ve got you’ve developed some things, some really, really cool stuff. And so can you explain to us in layman’s terms, because that’s the whole kind of core of what you guys do, too, is making AI accessible. So tell us about what you’re developing? What is your innovation?
Guillaume 11:51
Sure, I can start and certainly, Jonathan can begin. You know, I guess the best example is when I asked people, “are you a doctor?”, like Denis: Are you a doctor?
Denis 12:01
No.
Guillaume 12:02
No. Okay. Have you ever seen an MRI or an X-ray or scan?
Denis 12:07
I’ve seen an x-ray maybe once or twice in my life. And apparently, maybe on TV.
Unknown Speaker 12:12
Okay, fair enough. And Eunice, yourself, if you’ve seen these types of images? And you’re not a doctor, and neither am I but the beauty of the MRI or the X-ray is it gives the people that do research in medicine incredible information and insights into developing new cures, new ways of treating disease. But that same image gives you and me the ability to understand, you know, what is broken and what’s not working. And yet, we don’t need a medical degree, or the physiotherapist that’s going to treat your broken arm or your sprain. And so at Zetane, that’s what we wanted to do, what we said is that AI shouldn’t be left in the hands of solely the data scientists and the machine learning engineers, because if it is, it won’t ever really grow into a complete domain that is impacting the world around us. And if you don’t make it human-understandable, the way MRIs and scans are, then you can hold accountable the people that are giving you the solutions, either, right? Because you can’t say, “Well, you know, when you explain what you thought AI was, you weren’t far, both of you were pretty close. But the next usual question that we get is, “but how do we know it’s working?” How do we know it’s working properly? How do we know it’s not biased? How do we know when it’s not gonna work. And so what we did at Zetane, we said, we need to take this complex set of algorithms, [Python] libraries and code and make them visual so that not just the experts can go further with what they have. But the non-experts can participate, especially in industry, to validate that it’s going to work, it’s going to be tested, it’s going to solve whatever it has to solve, and it can be trusted and explained. So that is the thing, that is the software that we’ve developed.
Eunice 13:57
And it’s I love how you describe it, because at the end, it’s also about appropriation, it’s okay to have the proper solution for a very technical problems. But at the end, in the implementation of the solution, you really need a degree of appropriation at the completion, that it’s actually making lives easier. So on that what we know so far is that for many people, AI feels like a “mysterious black box”: you take some data in, some code, and then prove the outcome is for a machine that can help you solve problems. That’s basically the common understanding of this business. So as a person with no AI background, I understand that the value, the value of it, but not necessarily the application of the technology. And I agree with you guys, that shouldn’t be a criteria to exclude people from the discussion. So I think there’s a disconnection. So how do you help people like me to see the true potential and to understand the application of these technologies without having a real AI background or without becoming an AI expert.
Jon 15:07
Yes. So I think you mentioned kind of a perception of what AI is. And that, you know, there’s I think people think of AI as “smart” or even “dangerous,” right. And almost in a Skynet sense. I mean, we brought up Terminator earlier, right? But the problem is that models are actually quite dumb. They’re very — they’re focused on a very specific use case, in general. And they’re very — they’re limited in what they can do specifically. And so you know, that misconception is honestly, what leads to a lot of the challenges because when we’re talking about, you know, bringing out a model into production or understanding it, we actually, what we need to understand is what it actually can do. And what it can’t do. What it can’t do is a really important part, because that helps us build robust systems: it helps us you know, understand how it will fail, why it will fail, and what things we need to put in place to make sure that we can actually bring something into production and get it deployed. So, you know, visualization is a huge, huge piece of that. And Guillaume mentioned, you know, in the medical domain, you know, an AI model that’s predicting something, predicting a patient outcome, you need to — it needs to be verified with domain experts, with medical professionals. Not just about the outcome, but about what it’s doing internally, in order to make sure that it’s producing the correct outcome for the correct reasons. Because there’s a big difference between those two things, right? It could be producing a correct outcome that is trained to produce for all the wrong internal reasons. And that is possibly an even worse problem. Because we think that the AI model is working correctly, is doing what it’s supposed to, but it’s actually just completely “fragile.” And it’s going to be completely broken once it starts encountering the real world and unknown situations. So, yeah.
Denis 17:03
I think we’ve seen that and I know we talked about last episode as well, and ethics, because it creates some gaps, right? Like you teach it what you’re aware of, and what you’re looking at, or the data you’re feeding it and countless examples of, of a derailing going into completely the opposite direction, which we did not foresee at all. When it’s interesting, when we do talk about the kind of the potential future of AI, it’s, there’s so much more that it ends up doing. And then the things we teach it becoming much, much more autonomous in the sense, in that sense, but that’s where I spoke with your colleague, Jason and the conversation we had kind of pre-show to it when we were booking you guys. The concept of making AI more accessible and democratized is very, very interesting. That’s the word you guys use: “democratize.” So can you tell us a little bit about what do you mean by that? What does democratized AI look like for you?
Guillaume 17:59
Well, from my perspective, there’s two things. One, it’s back to the question of the black-box, right? Because if people don’t understand and trust something, they’ll fear it. And if you fear it, you won’t adopt it. And so, how do you break that fear cycle? You let people understand and explain what you’re trying to show them to a point where, you know, they’re comfortable with it. The problem is that, like, look at the example today of the game that we tried to play {referring to redacted section}. Let’s say that that had been an AI algorithm. It wasn’t, but let’s say that it had. Well, the risks of the consequence of it not working today were relatively minor; you know, like, we had a good laugh, and then, you know, you’ll learn and it’ll improve next time. But let’s say that had been a medical diagnostic; let’s say that had been, you know, an interview of hiring somebody; let’s say that it had been monitoring a car and its self-driving, you know, at 120 kilometers an hour on the highway, if it had failed. The risk and the consequence would have been different, right? And so that’s what democratizing AI in our speak is all about. We believe that if you can make AI human-understandable if you can do that, you put it in the hands of more people. If you put it in the hands of more people, more industry will use it and put it into the real world, like “our world”, and more of us will go “Okay, it’s safe. I understand it, and therefore I’ll use it.” And it’s no different than technologies that I’ve been introduced to ever since; you know, the last 20 years where at first were all like, a little stuck in the air, right? I remember, you know, the whole model, you know … the genomics field. You know, because people were able to explain it and give it into more, you know, generalizations. It scared them, you know, bejeezus out of a lot of people for many, many years. Until finally people start to realize, okay, there’s a lot of good that can come out of it. Same thing for us: democratizing for us is you need to have more people in industry, not just experts in AI. So developers, technical people, users, consumers of the AI, being able to interact with it in a way that opens up the black box, so they can ask all the questions — and Jonathan’s examples really good. And then say, okay, I can close the black box now, and I’ll trust the future black boxes, because you’ve shown me this one. But today, there’s still a lot of resistance in adopting AI because we haven’t opened that black box. And the demos you’re going to see today by Jon is good. I mean, they’re going to show you what opening, really opening an AI black-box.
Denis 20:30
And I’m personally really looking forward to that. And you know, that example of, for example, Tesla and the autopilot feature 100%, couldn’t agree with you more, because you can give your life to a machine to do what people have been doing for generations and generations and the decades — it’s kind of nuts to think of, “okay, I’m going to fully trust this thing to have life of myself, my passengers, other people on the road,” especially not seeing you know, how it’s analyzing what you’re, what it’s doing. You have no idea of, no way to kind of even predict what it might do in any specific moment. And that’s exactly where that’s true — that trust is broken. And how do you then interact with AI versus human that interaction is definitely damaged at that point?
Eunice 21:16
We kind of covered this question, but maybe Jon, do want to explain to us? What do you think is important? We understand that in the future, you’re going to rely a lot in AI-driven technologies or AI-driven solutions. But why do you think it’s important to you both, a decision to start working on this, like, sort of changing the discourse and start opening new options and new opportunities for people to jump in? What are the people, technologies and companies for you making this business for?
Jon 21:51
Yeah, I mean, I think, you know, AI, can AI and machine learning can do some really incredible things, in specific contexts, and specific use cases, right? And so big challenges, because of how much hype there is. And because of there’s kind of, you know, mixed terminology, maybe in the media of like, what AI is and what AI isn’t. It’s really important to build that understanding within an organization or within even an individual who’s in the field. You know, what are we actually talking about, like, what actually can AI do? And what can it actually not do. And that’s the only way that we’ll ever be able to deploy it. Because if it’s just kind of this generalized thing, this black box, or kind of a mystery, it will be impossible to make real progress in terms of not just in research. So you know, research is just continuing to sprint ahead in what can be done, and it’s incredible how much research is being produced every single day. But in terms of industry: I mean, there’s a real lag of that research crossing into industrial applications, or enterprise applications or small businesses, to get some actual value out of all this incredible research that’s coming out. And a big part of that is this communication process is actually very challenging. So between an expert and someone who may be an expert but in their business; the ability to understand, “okay, how can this technology actually help and not harm us?” “How can this technology actually bring some value to our organization to our customers?” that’s the piece that’s really been slow in developing. And that’s the piece that we’re actually very focused on. That’s the piece that’s important to us is kind of bringing that research, bringing that, that you know, whether it’s cutting edge or older research, but whatever can actually help an organization. That to me is the important piece. And that’s what we’re interested in.
Denis 23:45
Is there an example that stands out in your mind from the, you know, the partnerships you’ve kind of created, the work you’ve been doing that really shows, you know, the impact of what you’ve put together with the visualization that really kind of helps them break through the next next level?
Jon 24:03
Yeah, sure. I mean, Guillaume, you’re nodding — feel free to jump in. I may be thinking the same example; thinking of different examples…
Guillaume 24:11
Yeah, I’ll tell you. It’s really a good question. When we started Zetane we were focused on the software product and it’s launched; people can go to zetane.com to download it right now; there’s a free Viewer on top of it. So the product lives, right? But we were getting sucked into companies saying, “Can you help us for actual projects we’re trying to get done?” So we got brought into that field, and we do a lot of that work now. There has been at least three, four or five examples of us working with an AI team who had been struggling to get their projects approved internally by the business side of the companies with the decision-makers, the operational people that consume the AI that would come out of the innovation teams. And after we’ve helped them transition their work into Zetane. I’ve had — I was there — just before the pandemic hit — where they literally took their laptop, walked over to the VP of Engineering and the CEO — like those levels, C-suites, right? — and said, “I gotta show you something.” They showed them in three, four minutes, the Zetane view of their project they’ve been working on for months: immediate acceptance, not exaggerating. One day, like, I was there — it happened. And so we started realizing, like, this is really powerful, but people don’t know that it exists, right? But when you visualize something, and you’re told, like: here’s how the data is going to go through; here’s what the algorithm does — this fancy black box; here’s what it looks like; here’s why you can trust it because these are the things we’ve done; and here’s how accurate it is giving you that output, whatever that recommendation is. And you do that visually. And then if you can challenge me and say, “doesn’t make sense that: that part has so much impact while this part should have more impact,” you know? And you say, “oh, okay.” And the acceptance becomes a collaboration as opposed to me trying to convince you. So yeah, I could give you a bunch of examples, in computer vision, which we specialize in a lot. So that field is an exploding field, but also in regular tabular data, you know, like, “why should I trust?” “Well, I just showed you why — challenge me!” And if I can’t answer, I’ll come back to you with the answers, which you can do today with Python code and libraries and stuff that you can’t get your hands into, unless you’re an expert.
Denis 26:25
That sounds exactly like, you know, back in the day, when computers were just becoming a thing, and all you had was the DOS interface. And then Apple came up with, you know, the visual interface and [snaps fingers] — applications, people acceptance, like skyrocketed, because suddenly, you can use it for something; you can understand what to use it for it that makes a huge difference.
Guillaume 26:45
Yeah. Great example.
Eunice 26:49
What applications do you see — what revolutions of AI are you hoping to see, to be part of developing with the tools you’ve created?
Jon 26:59
So we, like you mentioned, we specialize in computer vision. Computer vision is an area that I think maybe has less adoption at this point. It’s the domain that is very interesting. And it’s also challenging: the tools that we’ve created that are visual, they help understand what computer vision models are doing, and how computer vision models work, specifically, and kind of the challenges and constraints around computer vision models. So it’s definitely kind of a passion of ours is working in that industry. We obviously — we do work with NLP, natural language and tabular data as well. But the tools that we’ve developed, the exploration and visualization tools, they kind of create a common language when we’re talking about a computer vision model. So that’s what we’re looking at here.
Denis 27:56
So we’re kind of coming up close to time for the demo. There’s enough time for maybe one more question. And I do want to let our audience know, to make sure that if you have a question for Jon or Guillaume, make sure you post it in the comments. Shortly, we’ll be opening up the mics as well. So you can join us on screen. If you want to, you know, leave a comment, say something specific about, you know, what we’ve been discussing and ask a question directly to our guests. So keep an eye on that. It’s definitely coming up, but before we jump into the demo, I do want to kind of touch on one little thing. And piece of that is one to congratulate you guys. I know at the end of 2020, you had some really big news with funding coming through. You got $350,000 to work on the Rheinmetall project, in part with MEI, the Quebec innovation branch of the government. But also I saw that you received $100,000 from the Canadian Economic Development, so far, huge congratulations on that. That’s very exciting news. But I’m really curious to kind of dive into this a little bit — and very briefly — but in terms of promoting innovation, from the government standpoint, how does input like this from the government make an impact on your ability to innovate and change the industry? And, you know, take your project and what you’re working on to the next level?
Guillaume 29:20
Well, I listen, you know, we’ve been able to get roughly $1.5 million of non-dilutive funding. So some of which you mentioned, and others, we’ve got some exciting press release coming out next week on a gaming company that we can’t tell you about yet, you know, granting us $150,000 for what we do. So, and so I would say that, on the one hand, you know, one of the biggest struggles I think that companies have in Quebec, in Canada, is having start-up technology get adopted within the government earlier rather than later. You know, by definition, governments are risk-averse. So they do a good job of putting money into programs that we’ve benefited from. And we wouldn’t be here today without Quebec, you know, provincial and federal and Montreal; PME West-Island helped us quite a bit as well, BDC. So we wouldn’t be here without them. So that part works relatively well. But the technology adoption piece of saying, “you know what? We’re going to use Zetane in the branch of government ADC because it’s going to help us do risk and understand and explain AI solutions. And yeah, it may not be perfect just yet, but it’s pretty good. So let’s jump in.” I think that’s the part that still needs to be developed. Whether it’s AI or med-tech, because I’ve worked a lot with med-techs or other technologies — I think that is a big part. If the government could do that, they would be outstanding, because the rest of the programs are pretty solid. I don’t know if that answers your question… Yeah?
Denis 30:57
Yeah, I think definitely, in part, it says that, like I said, it’s just touching the surface of the topic, because I know we have a lot of things still planned for the show. So maybe we can continue the conversation after the demo, because I’m very keen to jump into that. Jon, if you could share your screen. Let’s jump into it.
Jon 31:18
All right, let’s do it. Alright, can you see my screen? Yeah, okay. So what we’re looking at here is the Zetane Viewer. Like he mentioned, this is free for download on our website. And what this Viewer is looking at is an AI model. So you know, we talked a little bit about what an AI model is. But you know, at its heart, it’s a series of statistical and mathematical operations, that’s going to take input information and transform it into an output prediction. So in this case, you know, I’ve gone with a real-world use case that most of us have probably experienced. So most of us have probably interacted with a banking app, where you take a picture of a cheque and it uploads the cheque and puts the money into — hopefully — your account. This is a digit-recognition neural network. And so we’re looking at a smaller example of a digit-recognition neural network, that’s going to be taking in the input image that is produced by a user of the model. And it’s going to transform it into a prediction. You know, it could be a single numeric value, or a series of numeric values. And the model is doing, is processing that input information and transforming it into that output prediction. And so we’re actually we’re looking at the model input, right? Now we’re gonna look at the model; we’re gonna look at the internal information inside of the model, and then the output prediction that the model is making as it transforms that information. So in this case, the input is an image of a digit. And this digit comes from mNIST, which is a database of 60,000 handwritten images. So it’s designed for training models to recognize these handwritten digits.
Jon 33:29
You can see over here, this visualization of the input image: it is 28 by 28, so 28 pixels wide by 28 pixels high. And it’s coloured based on the values of each of the pixels. So this colour scheme, red is a higher value, and blue or purple is a lower value. You can as a human — it’s really easy to see that the ‘two,’ the drawn two, has the highest values. But that’s part of the visualization, is we’re actually showing the raw values. So this input is basically, you can think of it as a statistical distribution that is passed into the AI model. And the AI model is — it learns over time, based on making a prediction randomly at first, it gets corrected by a mathematical function called the loss function, to teach the model over and over again, whether the prediction that it was making is correct or incorrect, and get it closer to the correct output prediction. And so this, this happens, this learning process is called “model training”. And when the model trains it’s actually learning information about this input image, this input data, and it’s storing that information inside of its internal — what are called weights. These are — it’s learned patterns that it has recognized. And I can show that quickly. So, you know, we’re looking at the actual internals of the model here. This is a similar visualization, but these are the models weights. So this is the information that the model is learning about the input data that is passed into it. This visualization here that we’re seeing of the ‘two’, or this kind of coloured version of the ‘two’ is a combination of this particular input image and the learned weight information. So they’re combined in order to produce this visualization. And using that, you can kind of see what the model has learned about the number two. So this is, this computer vision model is learning things like vertical lines or horizontal lines; it learns about edges, curves; it learns about, like the juxtapositions between different strokes. And it also — one interesting thing you can kind of see here, so this is the same colour scheme. So redder is a higher value and blue is a lower value, you can actually see that there are parts of the ‘two’ that this particular weight has learned to avoid. Right, so when the model is rewarded or punished based on a prediction, it also learns the things that help it differentiate a ‘two’ from a ‘seven’, or a ‘two’ from a ‘three’. And so it remembers those things, and it is highly — it pays attention to them in order to differentiate that in the final output. So as the model passes its data through its operations, the input image actually is downsampled; or it gets smaller and smaller. So what happens is the weights, this model — it just, it’s the same process happening over and over and over again, as the image passes through and gets smaller and smaller, the weights actually start learning more and more particular features about this kind of image. And so you know, in a simple model like this, where the image is very small, it may be learning, you know, very particular things about what makes a ‘two’ a ‘two’; it might be just, you know, a group of pixels; maybe, you know, 10 pixels, that are influencing this model, in its prediction of a ‘two’ saying, “okay, you know, this is where the top of the ‘two’ meets the bottom of the ‘two’; that’s what makes this a ‘two’ and not a ‘seven’, or a ‘three’ or a ‘nine’. Eventually, these distributions, they’re transformed into a statistical likelihood prediction. And that prediction is going to be what the model thinks this input image actually is. I can show that real quick.
Jon 37:48
So in this case, it’s predicting one of 10 digits, so ‘zero’ is a digit here. And in this case, it’s the same colour scheme. So 0, 1, 2, this is the highest valued prediction. So the model is saying, of all these possibilities, ‘this’ is the most likely prediction of this input image. So that is basically a kind of a whirlwind tour of how a computer vision model works; they get larger and more complex than this, and you know, in a larger, more complex model, they start to learn things that are even more complicated. So, you know, a larger model may learn human recognizable features, like in the feature maps that I showed, you know, maybe the nose of a dog or an eye. It’s a very — it’s very common for computer vision models to learn about eyes for whatever reason; or you know, the rim of a car or something. These are all structures that the model will learn to differentiate these different kinds of input images. And, you know, based on kind of what we said earlier, you can probably tell that the model is not actually — it’s smart, but it’s not that smart. It’s doing one very particular thing. And it’s trying to do a very good job at that one particular thing. But the models can be, they can be tricked.
Jon 39:19
I can just show a quick example that I took last night of an incorrect prediction that this model made. So this is a digital input image that was passed into the model of a ‘nine’, and the model predicted that this was a ‘seven’. So we can look at, you know, the features that the model learned to predict which were diagonal lines, vertical lines, horizontal lines, it sees all of those. And you can see like, you know, the detection around the ‘nine’, but it didn’t do a very good job of learning the difference between a ‘seven’ and a ‘nine’ that’s written to look quite a bit like a ‘seven’. You know, just barely has this internal circle. And so the model actually predicts the likelihood that this is, you know, either a ‘seven’ or ‘nine’. But in this case, it actually thinks it’s more ‘seven’, it’s more seven-ish, than it is nine-ish. So, you know, this is one of those; this is just a simple example, and kind of a theoretical example. But in the real-world use case of cheque, you know, the image may have all kinds of challenges that need to be planned for and need to be mitigated. You know, maybe the cheque is not a picture of a cheque at all, maybe it’s too bright or too dark, or some key information is hidden from the cheque. And so even in a simple case, you know, the system is designed in order to get that model into actual production is about knowing what the model can do, and knowing what the model can’t do. And knowing how to plan for that, and how to work around that. So I can have more examples. I don’t know how much time we have.
Eunice 41:12
I think that is an awesome introduction, like thanks for walking us into this. Because sometimes, we rely on AI not knowing exactly how it works. But now, like, at least for the audience and me, it’ll be a bit easier, which are which are the limitations. Or what are we talking about when we’re talking about predictions and models based on either visuals or other tools. So, we are getting close to the interactive part of the show. So what’s next is to invite the audience to basically take the bank and make the questions for our guests. And yeah, whoever feels like jumping in.
Denis 41:57
Yeah, so if you want to join the conversation, one, you have the chatbox on the right, so feel free to write in there. And we’ll keep an eye on that and pose any questions directly to the Omen, Jon. Otherwise, you can click the microphone button, and then you’ll be able to actually join us on screen, which is a lot of fun as well; we would love to actually see you, who are you joining us today. Feel free to post some questions to our amazing guests. But in the meantime, while we wait: Jon, Guillaume, I was very curious. So seeing that visualizer, are there any kind of, like, limitations? What kind of AI applications can be visualized like that, or is that information is able to be turned into something visualized and usable for people who are not necessarily like AI-background minded? Or is it fairly applicable across the board that it’s open and possible to apply to almost any use scenario?
Jon 42:54
I mean, I think so. I’d say the strength here is computer vision models. The other example I have is a medical segmentation model, which, you know, if we don’t have any questions, we can dig into that one as well. Any situation that — the mNIST is, it’s kind of interesting, because any human, almost any human probably can look at a number and give a correct judgment of what that thing is, right? That’s a ‘seven’, like even if it’s poorly written as a human, like, we have a pretty easy time of differentiating a poorly written ‘seven’. I mean, maybe not in every case, right? Like, I don’t know, some handwriting is impossible to understand. But in more complex and higher-risk scenarios, like y’all mentioned, especially, you know, in the medical field, or industrial fields, the need to bring in domain specialists and other experts who are not just ML developers, and have that shared process of looking at what the model has learned, what patterns it’s learned to recognize in those input pieces of information, it becomes more and more important. So you know, specifically in the medical field: a model could be predicting something, but like I mentioned, it could be predicting it for the wrong reason. And so you need medical professionals working with ML developers working with AI specialists to share that process, that development process and visualization process in order to have that kind of back-and-forth and say, “okay, it’s working for the right reasons.” Or it’s not working. And this is why it needs to be fixed. And this is how it needs to be fixed. So you know, this, the visualization tooling gives kind of a shared language that people can interact with each other. That’s not specifically domain — you know, domain-specific terms, like doctors have a lot of them, ML developers have a lot of them, but visualization or images are placed where you can kind of meet and say, “okay, well, is that what I’m talking about? Is that thing that specifically is what I’m talking about when I say, you know, ‘pulmonary whatever’.” You know, I don’t know any medical terms. I try to make myself look too stupid.
Denis 44:58
Awesome. So I see we’ve got a couple of quick questions from the audience. I’m gonna scroll a little bit back up because I saw Gabrielle, you sent us a question a while ago, actually. So I’d like to touch on this one. So Gabrielle asks: “so obviously accessible AI should run with socially acceptable — accepted AI. So how would you go about getting the general public to understand AI enough to not uselessly block development out of fear?”
Jon 45:25
Yeah, I mean, that’s a tough question. I think it’s a tricky subject. So I mean, one of the things that’s important is, you know, there’s an issue with training information, and it relates to another question that was asked, but, you know, models, by definition are kind of dumb, right? They’re not really generalizable in the kind of general AI sense that I think some people think AI means or even are looking for, you know, coming out in the future, what is the general AI intelligence. In the case of models that are recognizing things, often, the problems with the data sets: so data sets that are incomplete or incorrect, or missing big sections of the real world is a big problem. With self-driving models is that in order to get enough data, to train a self-driving model, you often have to generate these, you have to simulate the data. And this can be problematic because your simulation may not represent the real world in any meaningful way. So even when you’re trying to increase the amount of data you have, you’re still kind of bound by the constraints of the developers; you know, how they perceive the world, how they think of the world, and the data that’s being fed into the model. So I mean, I think the, to me, the important piece of this is understanding basically the limitations, the limitations of data, and the limitations of development and kind of working in an organizational concept, or maybe in a social context with kind of a broader group of stakeholders, a broader group of people who are able to contribute to this development process. And those are the kinds of things that can, you know, move towards, you know, just more generalizable and better performing AI, in addition to, you know, what you might say, is more acceptable AI, or yeah,
Guillaume 47:16
I’m gonna add a couple of things, this is a really passionate topic with us. I’ll add two more things. You know, bias, and test and evaluation. So the problem with AI is it doesn’t know when it’s being biased, because it just takes data right and in the algorithm doesn’t get to choose the data it takes given the data. So if I trained an algorithm, and the algorithm for — exactly the way Jonathan said it — for what you created it, it’s really good at figuring out how to outsmart you. It really is. And so there have been plenty of examples where the algorithms has taken on one part of a data and said, “a ha, this is the driving factor that gives me the answer! So I’m going to cheat and forget all the other stuff, I’m just going to jump to the one that always seems to be the driving factor.” So if I’m giving you a population of data, which is definitely ethnically-biased, or gender-bias, then the outcome will go, “Ah. It’s always a guy. So give me a CV, it’s a guy, I’ll put them on top of the list. Give me a criminal of an ethnic background, they have put him or her on top of the list,” you see what I mean? And so the ability to open the black box and say, “where is it biased? what is it saying, really? Where is the data bias?” Because the algorithm in and of itself is not biased. It’s just smart at what it does. That’s number one. Number two is, I come from a world of modelling & simulation software; like, I’ve done a lot of that in my career. And in that body of knowledge, you know, we talked about software testing, software validation, software auditing, right, that whole and it’s been on for years. That doesn’t exist yet in the AI [sector]. So the concept of how do you test and validate AI to make sure it’s doing the right thing for the right reasons doesn’t exist. And one of the areas that we’re focusing on is really that part of the workflow to say, yeah, you can show your data you can see inside — like Jon showed — you said, “why are you calling this ‘seven’ versus a ‘nine’ or a ‘one’ versus a ‘seven’? That’s a difficult one, depending on the European one or a North American one, it can be a ‘seven’, right? For the four, it looks funny. And so what we’ve done is we’ve said we’re gonna make it visual so the bias can be exposed and people can have that conversation, have the data and the output. And we’re going to create an environment where you can do testing & evaluation, to address the question. Because until you do that, you’re trusting again, you’re trusting the [inaudible], you’re trusting that — just like you and I would trust a biased individual in a job. It’s no different unless you’ve trained that individual to look at the world differently. The data around the individual, it’s not going to be biased, right. So that’s what we’re trying to do.
Eunice 49:57
What did I can this rotation is at the end everything about adoption and they remember being at a nuclear facility in the US and they have like tons of regulations and tons of warnings on where to get close, we’re not to get close, that you’re going to use different kind of lights for to make more and more specific for people. You know, even if you speak the same language, it’ll be clear for you that you see a red light, don’t go there. But at the end, what I like the most is the sign for this legend saying, “machines don’t have brains; please use yours.” Yeah. So regardless, if it’s AI, or if it’s the printer from Gutenberg, or in a nuclear facility, at the end, we are the masters of the technology we are developing. And it’s important for us to recognize that it’s important for us to appropriate those technologies into grid to identify, as just mentioned, what is the limitation of the technology. And which are the potential avenues that we can explore with it. So it’s fascinating.
Denis 51:02
So we hear from Ella Andreena — I hope I’m pronouncing your name wrong, right. So to feed the data to recognize and learn may introduce a range of emulated bias to actually find the higher bias, is that correct?
Jon 51:18
But I shouldn’t have mentioned that I have a philosophy background. I mean, if I’m understanding the question correctly, which I may not be, I mean, it’s a very interesting point. And I think that can be a very interesting technique to discover — I mean, in many ways, you know, if the developers of the model, that when we’re talking about cross-functional, or like, you know, multiple stakeholders approaching the development of a model, it’s important to have representation of a lot of different viewpoints in the actual development, because the selection of the training data — the creating of the training data — can actually represent the biases that exist already in the developers of the models and those types of things. So when there’s certain groups of people who are the only people who are doing model development, then, you know, those biases consciously or unconsciously begin to get reflected in the systems that are created. So I think, yes, there’s absolutely a level of, you know, higher bias and the discovery of that from, you know, some of the large-scale failures we’ve seen in ML systems.
Guillaume 52:30
Yeah, I was gonna add, because I’ve worked on a lot of projects on consultant work. I’m overselling myself here — I’ve overlooked a lot of the projects and what I’ve realized is, you know, the famous data that you know, these data sets can be used well, by the way like, like “mega,” that’s what we call it big data. That’s what machine learning is so good at: the human will take too long to understand all the data and come up with an answer. So that’s the positive side. But these data sets have a lot of features in them. And what the machine learning algorithms and deep learning algorithms do is they figure out which of the features I can give you 1000 features. But really, it’s like, like the numbers that Jonathan was showing, there may be three or four key parts of the features that drive the majority of the results. And so you need different levels of bias and understanding of each feature has a predominant role. And when Jonathan was showing the model in the demo he showed you, you didn’t see that at a macro level. But what he was showing you is going left to right, how the data is flowing. And each along the way of those different sub-models, how each of those sub-models architectures, was having an impact on determining what was going on, and how it would classify it. And so the question that we were just asked, it’s understanding along the way, what happens. I’ll give you an example. We did some work with a company that was looking for bolder detection on a railway, railroad tracks, okay. And the pictures that they had used to train the model was a model — was a boulder just before the track ends entered a tunnel — boulder on a track just before a tunnel. If you take a picture of that, the boulder and the tunnel outline also looks like a boulder. The algorithm was actually telling the train to stop. So good decision based on the shape of the tunnel. It interpreted it as a boulder, not based on the shape of the boulder. And so that understanding — and we found that out after the 12 layers going left to right — that for some reason, it said: “aha, we’re going to stop because I’ve been identifying this shape as a boulder, which is this.” You know, there won’t be no boulder, he would have stopped for no reason, right? And so the bias is that his understanding along the way, what time did the algorithm say I’m putting more ‘weight’ here. So top level bias second level biases as you go ‘left to right’, as we call it in the model architecture.
Denis 54:51
So we’re running really close on time here and I want to make sure I give you guys a little bit of time at the end to let us know how to follow you but I do want to touch on one question here, because I don’t want to ignore our audience, Luke. So he, Luke asks, “does the same product allowed to visualize models not apply to 2d images? Like time series?”
Jon 55:12
So yes. I’ll give you an example. We worked on a, what’s called a wake word detection models. So when you say, “OK, Google or Alexa,” there’s a neural network that fires in those devices that’s specifically designed just to detect a wake word. It’s only purpose is hearing that one word, and then firing and turning the device on and going in calling home for the larger system to start interacting with it. So you know, sound is a time series: it’s a wave, right. But you can use a lot of different kinds of models. In our case, we use a computer vision model. And this was a particular model called ‘honk.’ It’s using a computer vision model to identify patterns and time series in the audio itself. And when you visualize it, it’s actually fascinating, because you can see what the wave pattern looks like. And you can start applying convolutional approaches, or visualizations to that audio waveform in order to see where the model is putting attention on the waveform. And so similarly, how you know, model can learn a pattern in image data, which is, you know, from the computer’s perspective, it’s just 794 unique values, there’s nothing image-ish about it — that’s just for our — the way we interpret information. Similarly, you can do the same thing with audio information, which is also just a series of values of different intensities. And so you can start pulling things like the ‘seven’ to ‘nine’ cases from the wake words, something we did. So we pulled, okay, here are the ones that are not performing very well. And you listen to them as a human, you say, “oh, okay, obviously.” I can tell them that’s not performing very well, because it sounds like nothing, or it’s random noise, you know. So there are many cases of the data being completely, you know, someone says, “go and stop,” but it sounds almost exactly the same, because the recording wasn’t very good. And the model is saying, “okay, well, you know, I don’t understand how to predict the difference between these two distributions.” So, yes, that is absolutely something you can do.
Eunice 57:21
So well as per tradition, we always wrap the Innovators show with nominations. So lets innovators know others innovators, whether mentors or someone, they look up to even rivals. So every show we like to dedicate a bit of time to hear from our guests. And I think the first question is, who’s the most innovative person or organization that you know, or that you’ve interacted, that you would like to nominate and hear from in future episodes?
Guillaume 58:11
Good question. Look, I think that I think mentored a few companies over the last little while I think that hearing companies like you know, Sarah Jenna out of My Intelligent Machines — a great success story in genomics here in Montreal — using AI and genomics, and a few other different things, I think, would offer female leadership, tech, Montreal in a pretty competitive environment. So these are some — that would be one person. You know, the work that Sylvain Karle is doing out of Founder Fuel now, you know, taking all his knowledge as the basically the guy who created founder fuel, and exporting it to environmental and socially responsible initiatives, so all that know how to build tech companies. But in that in that part, I mean, those are just two very different profiles that we put forward. But those are certainly two very, very, you know, interesting, people that would have a lot to share, and innovators in their own fields.
Denis 59:23
We’re right at 1 pm. So actually, I don’t want to I don’t want to hold you guys up too much longer. So but I do want to dedicate a quick segment to you. If our for our audience to follow you along, for whatever’s happening. This is kind of your time to give the call to action of you know, how do they follow you? What do you want them to keep an eye on? Take the next couple of minutes to let them know exactly that.
Guillaume 59:50
Well, on my side, I would have said go to our website, but I will say go to our website in about a week or so. Because we’re releasing a brand new website with just a brand new look, but also a clearer way of navigating and finding great information. So that’s a great way, whichever the old or the new website, the free Viewer is exactly that it’s free, it costs you nothing, there’s no information, you just download it. And you can start seeing what some of the stuff that Jonathan was showing you. They’re more aimed at machine learning, deep learning models, but you can really play around very, very intuitively. So that’ll be my second. And then the third one would be, for sure, if you’re in a company that’s looking to bring in AI, and you’ve been struggling to do so. You know, we’d love to have a conversation on that, for sure. So we can be found at Guillaume at zetane dot com, Jon at zetane dot com, or visit zetane.com.
Jon 1:00:44
Yeah, when one of the things — just real quick — we’re releasing is — we’re getting, we’re putting a curated set of models in the Viewer itself. So when you download the Viewer, there will be pre-curated models. So if you’re learning if you’re trying to understand what models can do, you can actually just use the ones that are already built into the Engine. So you there’s five of them of different kinds. So one is ResNet. Just classic architectures that you can kind of look at, and you can experiment with them. So you can come with some free sample data. And you can see how the sample data is transformed as it passes through the image. And you can load your own data and pass it through without having to you know, write a Python script or drop into code, just to kind of understand what’s going on and maybe to validate if you have custom models. So.
Eunice 1:01:33
Thank you, Jon. Amazing, thanks again. One more time, let’s virtually put together our hands and let’s thank our guests. It’s not in the chatbox, it just in the screen. And let’s show them our community love. So if you enjoyed this episode, please share this with all the AI enthusiasts you have in your life or your community. And over the coming weeks, we’ll be posting highlights on the Montreal NewTech social media, so make sure to follow us. And if you haven’t done it yet, to create you to follow us and to share and interact with us in social media. So Denis, it goes.
Denis 1:02:21
So we’re not really done yet. Do end up every show with a little bit of a breakout room. And Jon, Guillaume, I know you guys are very busy. If you have time, we’d love to see you in there. That’s a place for our audience to interact and have a bit of a conversation post show. So we’ll have two breakout rooms open one dedicated to continuing the conversation on AI. And the second one is what we call the VIP Innovators experience. It’s the behind the scenes of the show. So we always our main goal is to always iterate, do better next time. So we open up the doors to our audience to join us in that other breakout room to let us know, you know, how is the show for you? Did you enjoy yourself? What worked? What didn’t? I know we can have a bit of a conversation about the mess up today with the game for example. I’d love to hear if you guys would like us to try that again. So we definitely would love to hear from you over there. So otherwise, thank you so much for joining us. Guillaume, Jon, it was a fantastic conversation. And we’d love to continue that afterwards. If you have time or otherwise, we’ll definitely be in touch with you guys by email. So thank you so much for joining us today. Thanks for having us, everybody.
