The future of deep learning
Welcome back to What’s NEXT, a podcast from Samsung NEXT exploring the future of technology. In this episode, I talk with Yosi Taguri and Joe Salomon about how MissingLink.ai eliminates a lot of the grunt work associated with running machine learning experiments.
Ryan Lawler: Yosi, Joe, thanks for being on the podcast.
Yosi Taguri: Thanks for having us.
Joe Salomon: Thank you.
Ryan Lawler: To start, tell me, what is MissingLink?
Yosi Taguri: I’m Yosi Taguri, the head of MissingLink AI, which basically started as entrepreneur-in-residence project in Samsung NEXT Tel Aviv. Back at the time, I closed a company that did a lot of deep learning, which was pretty new three years ago. It’s a new technology that allows us to solve hard problems with data, basically, not with writing code. As an engineer, it really fascinated me that you don’t have to write a lot of code to solve really hard problems.
It was new, completely new. We started doing it, started practicing it, and noticed this is very hard, not so simple and you move very, very slow. So we set out to solve that problem of how can you build really smart machines really, really fast, like you’re building software today and how do you bridge this gap because we felt that something is missing. How do you connect, how do you link all those steps of building an AI machine from an engineer standpoint, not from a data scientist standpoint? It’s applicable also for data scientist.
What we want to basically do is we want to accelerate the pace of companies building really smart solutions to make our life easier.
Ryan Lawler: Yosi, talk me through, as a data scientist or engineer, when you have to set up an experiment and then rerun it, what are the challenges involved with that?
Yosi Taguri: First, you need data. That’s the new AP. Without data, you cannot do anything. With deep learning and AI, if you don’t have data, you have nothing. The algorithms are out there for free, but you actually need the data. So you need a lot of data and you need to understand what it means.
To make it simple, if I would build a machine that classifies dogs and cats, I would have a dataset, a lot of images of dogs and cats, and maybe some other things that are not dog and cats. For each image, I would tag it or label it with what it contains, a dog, a cat, or not a dog and cat. Then once I have that and it’s really ordered, it’s probably sitting in a folder somewhere, taking a lot of space as odd as it may seem it’s been managed like files in 2018. Even if you have a terabyte of files it’s still a shelved folder somewhere in the organization. Which is quite odd, because you would think that probably there is a database that can handle it and the answer is there is not. People are still doing file copying like the old Norton Commander days from the 90s.
Ryan Lawler: Okay, so what happens next?
Yosi Taguri: Then, you build something which is called a deep neural network, which is just a set of mathematical functions that are connected to one another. And what you do then is, you want that network to absorb the data and understand its meaning.
Now the way it’s done is basically you take some of the data, you understand what it’s meaning and you basically push it to the network. Then, you kind of ask the network, “What do you think it is?” It looks at an image and everything is completely random at first. The network tells you, “Oh, this is 50% dog.” You’re going to say to the network, “Well, I know it’s 100% of a dog. You need to tune yourself a bit.” So you do this step of tuning the network just a bit, not exactly, not fully going, “Oh, it’s going to be 51% of dog.” You do it again, and again, and again, not just with dogs, but also with cats and things that are not images of dogs and cats. Little by little, every step of this, every duration, this deep neural network starts to make sense of what is the difference between a cat and a dog.
Ryan Lawler: Okay. How does MissingLink fit into all of this?
Yosi Taguri: The problem is, when you’re running that experiment, you have no idea before running it if it’s going to be successful or not. How do you know if something is successful? It’s very easy. You take some of the data, you put it aside, we call it test data. You know what it means, you know if it’s cats or dogs, and when that training session ends, you take that model, and you just test it. You take those images, you understand what they mean, and you ask the model, “What do you think it is?” You know how good the model is now, but it really takes a lot of time to get to that point, in terms of- because it’s images, you need GPUs, it’s a lot of computing power, and time.
What happens if something doesn’t work? If it really doesn’t identify your test set like you expect it to do? Sometimes you have to change the wiring of the network. Sometimes you have to change your data. Maybe your data is not labeled correctly. Maybe you don’t have enough. You fix something and then you run another experiment, and another experiment, and another experiment, and eventually you need to keep track of what you did. What did you change over time? This is part of how you do good science, you document everything.
Basically, this is the first part of things, being able to track everything that you did. We want to make sure that nothing gets lost in the process. Sometimes people change a bit of code, run the experiment, they forget about the change, it’s not documented it’s not in their source control. Basically knowledge gets lost. So, what we do behind the scenes is we’re tracking every step of the way.
The second step is, when you’re dealing with data, sometimes you’re looking back at one of the experiments, sometimes you want to reproduce it. Sometimes you want to go back in time and actually do that step again. The problem is, even if you go to a database, any database today and you issue a query to that database, and that database returns an answer, that answer is just one point in time. If you will ask the same question again, anytime in the future, you are guaranteed to get probably a different answer. Why is that? Because databases are built to hold the current state of things, and they’re not able to give you like a time machine going back in time, what was my state three days ago? That is something that’s not built in any database.
While in data science you want to go back in time because you want to trace back and understand, “What exactly?” Which data points I used exactly at that point in time. We basically build a time machine for data. When you are asking our solution for data, we guarantee that no matter how data evolves over time, when you will ask the same question, you will get the exact result. The way we do that is similar to how you version code. Developers are always guaranteed to be able to go back in time, because they have commits, it’s version control.
The third part is, when you are doing those experiments, it needs to run somewhere. Where you’ve heard about GPUs, and a GPU is a special processing unit that is able to do a lot of math in parallel, and when you’re running and you have a lot of data, it’s really important to run on the fastest one because you want to save time, you want to get the results as much as possible. What it will allow you to do is basically, just completely automate all that scaling in the cloud of doing those experiments and tracking it. It’s on one hand experiments, you have the data, you have the hardware, and now you can launch as many experiments as you want and move as fast as you can.
Ryan Lawler: So Joe, what are the factors that are converging to make data science and machine learning so important right now? Why is everybody talking about machine learning?
Joe Salomon: It’s a combination of several things. First of all, there is a higher demand for complex scenarios today. Think of autonomous cars. This demand is being fed from the fact that we can actually utilize AI and the barriers that were, up til a few years ago, mainly were access data. This is the next revolution after the big data revolution. The big data revolution made it that almost any entity in the world has access to huge chunks of data.
And the next spot is, of course, compute. Deep learning as a technology, exist for over 30 years, but only in the last few years, the conditions were met. So, almost any company, not just giants like Google and AW and Amazon and Facebook can do it, but actually any startup with a cloud account can now run it. The key is actually getting access to the data, and not just the data itself, but I would say, curated data, the tag data, this becomes a question of how do I clean the data? And, how do I add more knowledge on top of the data, and tune the data to the problem that I’m trying to solve. Not just putting my hands on data. Of course, this is the basic requirement, but it’s not enough. This is key when you want to solve problems with AI.
Ryan Lawler: Right. So one of the things that I wonder is, it seems like there are a lot of companies that are able to now run these experiments, but they’re kind of running these experiments in silos and there’s probably a lot of duplication of people working on the same problems. What’s happening from that standpoint, and how do we make sure that if you’re one start up in Israel, you’re not basically working on the same dataset and trying to solve the same problems as a company in San Francisco? It seems like there would be a lot of wasted resources in that type of scenario.
Yosi Taguri: You know, innovation happens in parallel in many, many places. Smart people think of the same problems they want to solve, and they approach it in different ways. Even in Israel there are many, many companies that are doing medical detection, we work with some of them and- First they have different datasets, right? They get to solve the same problem in different ways, so five years from now maybe we’ll be able to converge those different styles into one thing, but it’s not necessarily bad, we actually want as many smart people that have access to data, to try to solve as many problems as possible.
Joe Salomon: I would look at it from another point of view. I would say that you want the competition, you want to create, because the question is not that this effort is going to waste, but because the insight of what’s going to work and what’s not going to work is not trivial. It’s not that you can look at the problem and say, “Oh, sure, this is going to be solved like that.”
Still, the domain is so young that the insights are still being built, so maybe, as Yosi mentioned, a few years from now it will be redundant to some extent, but today we’re still in the pioneering days of AI. Well, you don’t know what will work, and then you need all those teams to compete against each other. Very similar to what’s going on in life science. A lot of researchers are trying to find a cure for cancer, or different types of cancer. They don’t know which one is the leading research team. The same goes for AI.
Ryan Lawler: Well, even if we’re talking about different teams approaching the same problems and coming to similar solutions, or perhaps different solutions. Still, there’s this problem of, there’s a certain amount of infrastructure that each of them needs to set up. So, I imagine all of these teams kind of reinventing the wheel separate from one another. Is that the case?
Yosi Taguri: First of all, you definitely need infrastructure for everything. Now, when you build a new company, you don’t think about it, but you’re going with the cloud by default. It wasn’t the case, like seven, eight years ago. So the big companies like Uber, and Facebook, and Google, who really understands the value of data. It’s not by mistake that they’re collecting so much data, because they need to process it. When they approach this need to process it and actually learn from it they ended up building platforms that allows any engineer on the team, on the company to do data science, to do deep learning, to do machine learning at scale, so they can all be productive and make the most of the data and talent that they have.
When a new company starts, you know, a 30 person company, it’s not going to build infrastructure for doing that, that would be a waste of time. Even Facebook is not built on day one, although they had the data. It took them a lot of time to get to a point where they understand, they need to build something around this.
So, what we are saying is that, companies who are very concerned about their engineering health, would start to build little tools inside the company to make their lives easier. It’s like the lazy developers movement. They want to sleep, they want to go early back home, and they want some kind of an automatic tool to take care of things.
This is where we come in. We’re basically saying, “We want to save you time.” You shouldn’t do the same repetitive thing every day. It doesn’t make sense, cause you’re probably a data science that gets paid a lot, and you’re still doing really boring things like DevOps, like spinning machines, like copying files. You want them to focus on data science, not on DevOps. That’s where we come in. We’re saying, there’s such a thing called deep ops, which is deep learning operations, and we’re all completely automating this for you.
Ryan Lawler: Okay. So, let’s talk about your customers. Who are they? And how are they using MissingLink right now?
Yosi Taguri: One of them is AI doc, a very exciting company that scans- CT scans. Basically they allow every doctor to become an expert doctor in his field, because he has this access to a support system that doles it attention to things that are really early on like, doctors might not look at. So, they can look on a pixel and tell you, “This is 80% chance that five years from now this is going to be a tumor. You have to take a look at that.” And just imagine, what is a C-scan, right? You get into that machine, and basically the machine takes images of slices of your brain, right? And it takes dozens of those, so the human brain, even the doctors, they cannot handle that amount of data. So, what do they do? They basically throw it away and just focus on a few.
Now all that data, amazing data get thrown away. Nobody’s looking at it, it might hold some indication on what the future holds for that patient. What they are doing is basically taking all the data and making sense of it, because they have historic data from a few hospitals, from a lot, actually a lot of data, they’re able to infer, how does the future looks for you when you’re handing them your CT-scan. It’s quite amazing, because it’s becoming affordable.
Ryan Lawler: You mention them throwing away this data. Are they actually just deleting that data, throwing away, or are they storing it somewhere and not looking at it?
Joe Salomon: It’s the later, today you store all the data, but the thing is that the humans cannot possess that amount of data, so basically you take a subset of the data, usually a very small subset of data, and you base your analysis on that. That’s the beauty where a machine can access all, the entire data. And because a machine can process much more, and in parallel, they can reach a conclusion that might be where humans would make an error. So, one of the promises of AI that in some cases, they will surpass the humans. For example, even today, if we are talking about imagery cognition, so AI today is doing better job than what human is doing in the same data set.
Ryan Lawler: So, you mentioned AI doc is one of your customers, what are some other use cases?
Yosi Taguri: One of the interesting companies we have in Israel is called Nanit. Nanit has a smart baby monitor, so basically they position a camera inside the babies cradle, and they started a smart thing about it, if the baby wakes up, they’re able to tell you he needs help, or let him cry for a few minutes he’s going to go back to sleep. That’s what we think, cause you know, don’t have to attend to it all the time. It’s basically a security camera for babies, and very, very, successful, and we’re able to help them scale the operation in trying to build a better model, and better camera, and better system, that predicts for the parents.
Ryan Lawler: What does the future look like if MissingLink becomes ubiquitous?
Yosi Taguri: What is the future for humanity, with relates to AI right? I highly commend reading the book, “The Post-Human Series.” It really will blow your mind about what can be done with- Where are we going? Obviously, we’re going to- It’s in the far future, but we’re going to merge with machines somehow, with nanotechnology empowering you to be faster, stronger, live longer, and such. In the short term I think, what would happen is that, we would live much longer, not thanks to medicine, but thanks to early detection. To preventing you from doing stupid things, right? I think it’s quite obvious that nobody would die from car accidents if everything were autonomous cars, right? The highest risk factor are humans, we are going to eliminate that. As a species we’re going to take away all the things that kills us early on. I think we’re going to have a lot of things get done for us, in a completely seamless way that we do not expect, right?
You live in the future in America. You have Alexa’s and you can order things like toilet paper, and after 20 minutes somebody would knock on your door and hand it over to you. This is magical to me, we don’t have that in Israel. But this is, every kid in America, well Alexa, they don’t think about it anymore. Five years ago it sounds like science fiction, “What do you mean I could press a dash button and get, like anything I want? This is insane.” So, we’re going to see more, and more of that at an accelerated pace.
Some say, a lot of jobs are going to get extinct, right? And the same happened when the spreadsheet was invented, right? There’s an interesting Planet Money podcast about it. But, guess what, 10 times more jobs were created to deal with spreadsheets. So, we are just getting on another cycle with a revolution, where there won’t be any taxi drivers 50 years from now period. Probably no drivers at all, but there’s gonna be a lot of other jobs that we don’t even know how to define them now.
I have a kid, my smallest one is six. I have no idea what she’s going to do, like 20 years from now. Maybe it wasn’t invented yet. So I think that we are going to have most of our lives completely automated. We’ll have more need for smarter brains, so I think that we’re- It’s not by mistake that we think our kids are much smarter than us, they are actually much smarter because they have access to technology that drives them much further than we could when we were kids. We’re basically going to live longer, be smarter, and we’re gonna have much more jobs than we have today, moving forward.
Ryan Lawler: Okay. That’s very ambitious. If you weren’t doing this, what technology would you be working on? What would you be interested in?
Joe Salomon: I would say that a company like AI doc, for example, is a company when I’m looking at, and I’m saying, “Oh, well they are actually doing something that if they’ll manage to do it, it’s amazing.” You can hardly say in the high tech industry that you’re actually saving lives, right? You do some things, at the end of the day it translates into dollars, right? And, this is the first, well not the first, but one of those companies you can say, yes, if they will manage to fulfill their promise, they will actually save lives in a way that they will say, “You know this guy? I saved his life.” And for me, this is like, brilliant karma and if we won’t build them, and our goal is not to support one initiative like that, but to support thousands, or tens of thousands of initiatives like that. That’s what draws me into MissingLink, but if not, then I would try to build a product like this. Because, there’s a huge opportunity around AI.
Ryan Lawler: Last question, what’s one controversial opinion you have, that’s really strongly held?
Yosi Taguri: I’m really passionate about optimizing my time and optimizing workflows, especially around engineering. I really hate doing the same thing again, and again, and again. Through all my career I always automated things. It’s a- It really bugs me that manual QA, quality assurance is not reusable. It’s like you take a human beings, they have this amazing machine, it’s a brain, and you’re making it do the same thing again, and again, and again.
So, one controversial thought I have, is I don’t think QA has a right to live. I don’t think they have a right to exist. I think we can completely automate those, like in any company that I had in the past 10 years, I never had a QA, because we managed to automate everything. And guess what, we ship better, more secured, more stable products, because I basically made QA reusable, which you cannot make with humans, right? With QA, I think it’s a waste of time and money, and when you want to build a real serious solution, humans must not be involved with QA. So, this is one controversial thought I always- keeps bugging me. We don’t need them. And it sounds a bit, right? It sounds a bit hard, it’s people’s lives, so anything that can be repeated by humans can be accelerated by computers, period.
Ryan Lawler: Well, I mean, this is a question I have, just in terms of people talking about, kids learning to code, in 15, 20 years, will there be a need for developers and engineers in the same way, or will computers be able to write reusable code smarter than they can?
Yosi Taguri: Yeah. So on the same note, I think that what we’re doing today is basically making sure that 20 years from now we will be out of job, as engineers, as developers. You will need less of us to build more powerful machines, because we will automate most of it. But it means, you know, the world will be a better place. People like me, in 20 years from now will do something completely different, and it’s not necessarily bad, but definitely, people will write less code because of AI. We know that today.
Joe Salomon: I think about it a bit differently, or from a different point of view. Very similar to the fact that today you write less code than what you wrote 30 years ago, but you have orders of magnitude of people who actually in that industry. So, I would say that building systems, and building solution would look different when you have different skillsets. You wouldn’t write code as we write code today, but you write much more code.
I would say that the full stack engineer of 10 years from now will have a different tool set, probably AI will be a permanent part of the tool set, but you will do code differently, but they will code. It will just be done different from the Fs and Ls and loops that we do today. It will be a much broader definition or attributes, but it will mean that the time to market of new initiatives, or new system will be much faster. Very similar to- Think of what it took to create a new website, when the internet started, like 20 years ago, right? It would take months for a team of, like 10 people to create one single website. And today, we do two to three every day, a 12 years old can do it in several hours, so that will be the future of development. I would think it would open the ranks for many more developers than we have today. Just not coding as we think of it, like we’re doing today.
Ryan Lawler: Well, Yosi, Joe, thank you both for being with us for the podcast.
Yosi Taguri: Thank you very much.
Joe Salomon: Thank you very much.