Welcome to episode #029 of the Super Data Science Podcast. Here we go!
Today's guest is Chief Data Scientist Ben Taylor
We are joined today by a highly accomplished Deep Learning expert and enthusiast. Tune in to hear Ben Taylor share his well thought-out views on Artificial Intelligence and its place in the world of our future.
Ben also offers valuable advice and experience as a startup co-founder about building a startup in the space of Machine Learning in these exciting times.
This is a truly exciting episode – let’s dive in!
In this episode you will learn:
- Why Recruitment is Ready for AI Disruption (14:36)
- Advice for Anyone Looking to Build a Startup in Data Science/Machine Learning/Artificial Intelligence (27:24)
- A Crash Course in Deep Learning (30:11)
- Things that are Difficult for Computers (44:42)
- Things Machines can do Better than Humans (48:18)
- How Much of a Threat is AI to Humans? (54:10)
Items mentioned in this podcast:
- Ben’s video with coding examples
- Sam Harris TED Talk “Can we build AI without losing control over it?”
- Hard Work Won’t Make You a Data Scientist
- Data Science Type E
- The Lean Startup by Eric Ries
Kirill: This is episode number 29, with Chief Data Scientist Ben Taylor.
(background music plays)
Welcome to the SuperDataScience podcast. My name is Kirill Eremenko, data science coach and lifestyle entrepreneur. And each week we bring you inspiring people and ideas to help you build your successful career in data science. Thanks for being here today and now let’s make the complex simple.
(background music plays)
Hello everybody and welcome to the SuperDataScience podcast. I hope you're having an amazing day and today we've got an incredible guest. Today we've got one of the most inspiring thought leaders in the space of data science, Ben Taylor. Now Ben is a speaker, so you might have seen or heard him at a conference, and also Ben is an author of multiple very insightful, very profound and even sometimes philosophical articles on the topic of data science on LinkedIn.
So what you need to know about Ben is that he started off as a chemical engineer and he's done a lot of great work in that space, he even has a patent in nanophysics. Then he moved on to being a quant in a trading firm, where he applied data science methods and worked with GPUs, Hadoop, and so on for the purposes of trading financial securities. And then he moved on with his career, and now he is redefining recruiting via artificial intelligence. And the things that he's doing there are just incredible, but you'll hear from the podcast how they use video recognition to assess a candidate's performance during an interview and really speed up the recruiting process for so many firms. It's incredible how machine learning algorithms can be used to redefine even something like recruiting.
In this podcast, you will learn lots of valuable insights, Ben will give us a crash course into deep learning, and give you an intuitive understanding of deep learning and why it is superior to other types of machine learning algorithms.
Ben will also give you some tips on how to become an A-class data scientist and what you can do to improve your career, and much, much more. Does that sound exciting? I bet you can't wait to get started, so without further ado, I bring to you Ben Taylor, Chief Data Scientist of HireVue.
(background music plays)
Hello everybody, and welcome to the SuperDataScience podcast. I've got a super special guest, I've got Ben Taylor here on the line from Utah. I've been a big fan of Ben's for a very long time, and finally I've gotten him onto the podcast. Hi Ben, how are you today?
Ben: Good. Thanks for having me, I'm excited to be here.
Kirill: I'm super excited about this opportunity. I've been following your posts on LinkedIn and just your career for over a year now, and you were a very big inspiration for me, and so if people haven't heard of Ben, Ben is a huge thought leader in the space of data science. He's got some great articles, you may have read his Data Scientist Type E article, very, very interesting, very inspiring. And today, probably we'll get started off – well actually, just before the call, we were talking about Ben's background. As I told Ben, I wish I had recorded all of that during our interview, because that was the first time we met. But if you don't mind, Ben, could you please walk us through a bit of your background again. How did you get started into data science, and how did you end up in the position you're in now?
Ben: Yeah. So I think for most people who have been professional data scientists for the last five years and longer, they had no formal programs before. And so a lot of their backgrounds are very mixed. They come from physics, biostats, chemical engineering, anywhere really. It doesn't matter what their background is. And for me, it was chemical engineering.
So I studied chemical engineering because it was a good technical route that could still allow me to go into medical school if I wanted to become a doctor. So for people that don't know what chemical engineers do, their typical jobs might be something like designing a chemical process for a petroleum refinery, or a lot of these semiconductor plants, like Nvidia and Intel and Micron, they hire chemical engineers as well because it's a chemical process.
So I did that for my undergrad, and I was really lucky, because during my undergrad, I had my first real exposure to some serious machine learning programming. It was a numerical methods class, and it was an introduction class, but our teacher exposed us to the base introductions that you would expect, but we also learned things like simulated annealing, and generic algorithms that I would not expect to learn in an entry level class like that. So that teacher did me a huge favour to kind of inspire me on this concept of your computers working for you while you're away, or while you're sleeping. So I just love the idea that my computer at home, or my cluster at home, could be doing significant work 24/7.
And so for all of my engineering projects, I always brought in a machine learning component, or a machine vision component, even if it was a huge stretch. And a lot of the times it was. And in some cases, it even frustrated my chemical engineering professors because they saw the stuff I was doing as being kind of too far away from the core discipline. One example of that is I went and did a – we all did internships our junior year I think, that's the year before you graduate, and a lot of the other students went and did distillation columns, and worked for chemical processing plants. And I went and worked for the Desert Research Institute doing satellite image processing, where we'd pull in satellite data and we'd predict algae content in the ocean by combining multiple satellite image sets. And it involves a lot of programming and image processing, where they felt like that had absolutely nothing to do with chemical engineering.
But I enjoyed that, and I went and for my Masters, I moved from Reno, Nevada, to Utah and I was studying gold nanoparticle arrays, and for people that aren't familiar with those, they're exciting because they can be used for single antigen detection, so if there's a single virus or a single cancer cell, right now with today's methods, you can't detect it. It needs to be significant. But with these gold nanoparticle arrays, they're so sensitive they can literally respond to a single antigen. So a single virus could be detected, or a single cancer cell. And for all of my Masters work, since I had a strong programming background, I ended up doing all the machine segmentation, image processing, on these same images, scanning electron microscopy images, from those gold nanoparticle rays.
And then I went and got my first real job working for Intel and Micron in their main NAND flash producing plant. So for the Intel SSDs, all of their NAND flash memory comes from this plant that I worked at. So I worked there for 5 years doing process control and yield prediction, still chemical engineering things, but kind of had a statistical side to it. But the big breakthrough for me that I am grateful for, but I'll never doing again, I had an opportunity to go and work for a hedge fund as a quant. And a quant is slang for a quantitative analyst. And the interesting thing about quants is quants really are data scientists, and they've existed long before data science was a thing. So quants have been around for the last 20 or 30 years. And what a quant does is they can program really well, and they also understand ridiculous amounts of math because they do algorithmic trading. So they write programs that trade stock automatically. And the majority of all stock trades now are automated.
So this hedge fund, it was focused on trading on the news, so using sentiment. So all of the news articles, blogs, anything that could affect price for the top 1500 stocks in the US, they were analyzed in real time and stock decisions were made at this hedge fund. The interesting thing about the hedge fund, we built a 600 GPU cluster in house, and that was 4 years ago, and even today, that would be very ambitious, and 4 years ago, that was ridiculous. But we did it. We succeeded. We built it and it would do over 10 million 5 year back tests per day. It allowed us to do a lot of optimization.
And then I went back. So if you could see my face, or if you've seen a recent photo of me, I'm getting some grey in my chin. And the joke is all of that came from the hedge fund. I did get literally – within a few months, my wife noticed that "Man, you're getting a lot of grey quickly!" And the reason for that was, that working in a hedge fund, if you've talked to people that have, or if you have yourself, it can be very stressful, and there's a lot of money to be made, the people you work with are extremely talented, very smart people, in some cases abrasive, and when I would go into work, I would feel like I was going into a fight. Sometimes work, you kind of had that heightened sense of – it's hard to explain. Just the feeling you get before you have a confrontation. Imagine having that first thing when you come into work and it lasting all day, and then just that's your life. That's how you work. And that's not really a work environment that I could – I learned a lot, but after a year, I went back to Intel Micron. And then after a year, I realised I wasn't happy there because I wasn't feeling challenged.
And there was a startup called HireVue, they were backed by Sequoia Capital, they were looking to bring on machine learning, and so I joined them three years ago, and I helped to build out their data science team, their IP, and their machine learning component. Real quick, what HireVue does is they do digital interviewing, so we're all familiar with – everyone listening has had to interview, we've had to get a job. So we've gone – a traditional job will be, you submit a resume, someone calls you back, and maybe do an on-site interview, or a phone screen interview. Then that leads to an on-site, and then you get a job. Where with HireVue, they do interviewing right on your iPhone or your laptop. So you can go straight to interview. So for HireVue, for all of our data science openings that we have, anyone can go straight to interview. Where they do the full coding assessment, they do the interview, and then we've had candidates come through that process where I would feel comfortable giving them a job just from that. Typically we will do a live interview after that. But the interview process saves employers a lot of time, and half of the Fortune 100 companies use us. So that's Chase, Goldman Sachs, Unilever, Red Bull, IBM. A lot of huge people use us to do their hiring. And then the other thing I’m doing right now is I’m also the Chief Data Officer for a new startup called Ziff, and it’s a deep learning startup. So I’m doing both of them simultaneously.
Kirill: That’s really cool. And yeah, it’s very interesting, what you mentioned about the hiring process. Could you give us a bit more detail about what we talked about earlier, the way you use machine learning and artificial intelligence in the hiring process? I thought that was just fantastic and phenomenal.
Ben: Yeah. So one of the things I like about HR is everyone can relate to it, but it’s also extremely personal. So if you’re building models to predict yield on a wafer or which marketing lead to follow up with or something like that, it’s not really impacting anyone’s life in a big way. But if you’re writing algorithms that predict if you should get a job or not, it’s a personal thing and it could have a good influence on people and it can also have a negative influence.
So when I first joined I saw a lot of value in all of these digital video recordings that were saved in the cloud and what we ended up doing is we enabled machine learning on the video interview itself. So, from a 20-minute interview, we predict whether or not you should get the job. And what that allows is it allows for companies with really high volume to interview in the thousands or tens of thousands. We even have some companies that interview hundreds of thousands of interviews a year. They now have an ability to bring the top talent up to the top of the queue and react to them very quickly and in some cases advance them on through the hiring process.
We think we’re starting to see opportunities now where they could be hired just using the machine learning from the video. And the features that we use—we use voice detect, so using deep learning we transcribe the interview to get all the content, so all of your language can be used to make a decision. And then we also use micro expressions.
There’s an old show called “Lie to Me” – Paul Ekman, that person is actually a real person. We got to see him in San Francisco. He’s an expert on facial micro expressions and deciding what they mean. So the computer will actually measure micro expressions during the video, so that can pick up on a lot of the soft competencies that are missing on a resume or even on a phone screen. And all of that is brought together in a holistic model that makes a prediction.
Kirill: Yeah. And you said you can do a much better job than a human can.
Ben: Yeah. So, the fun thing about that point is, I’ve gone to different conferences to speak on this topic and show results and talk about case studies and stuff. And initially there’s a lot of disbelief because I don’t know if the technology needed seems like it would be intimidating or something that we’re not ready for in 2017. I don’t know why people hesitate that this is a possibility because I think it seems pretty obvious that of course this would work and it would work well.
But the point I’d like to make is humans set the bar so low when it comes to hiring and recruiting that it was actually trivial for the computer to step over it. And for people who aren’t familiar with recruiting, if you think about it, there is a huge luck component for everyone. If you want to get a job at Google, if you want to get a job somewhere else, depending on their process for bringing people on board, there is a luck component where if you talk to this recruiter versus that recruiter, you may or may not get through. The other thing that humans usually have is there is a similarity bias. I’m more likely to hire someone more similar to myself than I am to hire someone who’s a polar opposite. And then the biggest fault of humans is we can’t compete with a computer. And this is just true in general, it’s not just hiring. So, machine learning versus the human – the human in a lot of cases has a very difficult time competing with just observation count.
So a human might experience five hires, ten hires, a hundred hires in their lifetime, but a computer can comprehend 100,000. And then the other job against the human mindset would be even if they were looking at the exact same dataset, they both have access to 100 hires, the human will overweight and underweight experiences unfairly. The example of that would be community college. Let’s say I’m interviewing a data scientist or an engineer. They go to this particular community college. They come and interview and they are a complete train wreck. Not only are they a train wreck, they’re dishonest, they are the worst interview I’ve seen all year. So I will take that information and the next time I get someone coming from that community college, I already have a bias against them and they may not even get an interview now.
When we talk about it we realize that that is totally unfair. But humans do that all the time and they do it on the flipside where there’s an Ivy League bias. So if you went to Stanford or Harvard, you may get an unfair Ivy League bias that would suggest that you’re better than you really are. A computer is much less likely to do that. A computer is not going to say, because you had one idiot from this college or one genius, that that is meaningful. If you get enough, it will begin to realize that, but it won’t overreact.
Kirill: Yeah. And also, humans probably tend to have a recency bias that if you had one really crappy interview just recently and then you have a moderate one next, the moderate one compared to the most recent one looks like an amazing person but in the overall sample it might be just a moderate interview.
Ben: Yeah. It’s amazing to me, when you look at some of these companies and how they do their hiring and recruiting, how terrible they are. My first interview—I didn’t know if I was going to go to graduate school, so I interviewed at a nanobattery technology company, some nano-lithium battery company, and I had six interviews that day. So I interviewed with one manager, went to interview with another manager and another manager. So not only did they waste my entire day, I wasted—if you look at the man hours and the opportunity costs, that was six people, and there was a ton of redundancy in the questioning. They all asked me very similar questions. And the very first person I talked to, I made a technical mistake in the interview and they corrected me and they told me what the answer was and in the very next interview, they asked me the same technical question and of course I got it right. And in the third interview, they asked me the same type of question and of course I got it right. The problem with that is they have different experiences with me where now, when they sync up that evening after I’ve left, half the room could think I’m an idiot and the other half of the room thinks I’m exceptional. So I ended up getting a job offer from them which I turned down and I went to graduate school. But just that process is such a mess, and it’s actually more common than you think with respected companies where one candidate will interview multiple times rather than having a structured interview that’s done once.
Kirill: Yeah, that’s really cool. And it sounds awesome, this whole idea of interviewing through machine learning. And as I imagine, the company you’re working for – HireVue – is pioneering this space, and you in particular are creating all these algorithms and right now they’re available to big companies who hire thousands and thousands of people. But when do you think this kind of concept is going to become democratized, that it’s going to be available to anybody, to any small firm, or to any startup to apply your algorithms or approach you guys to assist them in the process of hiring. Do you think that will happen any time soon?
Ben: For HireVue specific as a company, they really are the enterprise choice in their space. So they kind of cater toward that size of a company. But for smaller companies, there are so many machine learning solutions that are creeping into HR right now, whether they’re doing resume predictive modelling or audio only. There’s a lot of companies that do these predictive assessments where you answer 50 or 100 questions and they predict outcomes from that. They can even accept open-ended responses. So it doesn’t matter how big the company is, there are definitely machine learning solutions that can help the current process. A resume model is probably the simplest one, assuming that they have enough volume to find this correlation. The volume definitely helps. I think it’s the future. I would be very surprised if my children, who are under 7, I’d be very surprised if they’re interviewing with humans when they are adults because of the benefits that come from what a computer can see.
Kirill: Yeah, fantastic. That’s very, very solid view of the future. All right, what about your second startup? You mentioned Ziff, where you’re doing some deep learning. Can you tell us a bit more about that?
Ben: There’s always a side project going on with me. There always has been and that’s changed depending on the year and what’s going on. But with this Ziff, I’m diverting a lot of time into it now and I’ve got a great co-founder who’s also—I see him as being a rock star data scientist who has a lot more business experience than I do. His name is David Gonzalez. So what we’re doing is we’re trying to lower the barrier to entry as low as possible. And when you look at these bigger offerings, Google has a predictive Cloud option, Microsoft does, Amazon does, IBM does. One of the issues that we’re seeing -- these companies are fantastic. They all have their own strengths. Microsoft is exceptional when it comes to integrations. They can integrate with Excel and all these different databases. It’s a very nice move if you’re a Microsoft [indecipherable 21:52] to kind of bring these solutions in.
IBM have been doing a lot of good for the data science community with Watson and the marketing, getting the message out. A lot more people are familiar with data science and machine learning because of IBM and because of some of these wins that Watson has had. Personally I really enjoy Amazon. I feel like they do a good job with the documentation. Google is a popular favourite, but they do some things well and they struggle with other things and I feel like their machine learning offering is something that they have struggled with historically but I think they’re improving with that.
So, IBM, Microsoft and Google have pre-trained models. I think even Amazon has some coming online now where if you want to predict age or gender or whether or not an image is safe for work, those are available to you know, but they actually don’t have a lot of models. They probably have less than 10 when it comes to image classification, but we really want to blow that up with the marketplace. So I’m really excited about that. That’s what I’m working on now, but I’m still actively involved with HireVue and working with their team and their roadmap.
Kirill: Very nice. So you guys are going to be looking for funding, or are you just going to be doing it all on your own?
Ben: It’s interesting because we have talked to VCs, we do have funding opportunities, but it’s always kind of a dance between you and the VCs on do you want to take the evaluation that they’re willing to give you, or do you want to hold out for the evaluation that you want. Luckily for us, right now we have enough revenue coming in that we don’t need funding at the moment, but funding of course leads to some great things that would help us like bringing some key players on board sooner. So I think the plan right now is hopefully soon during the next couple of weeks you’ll start seeing some press releases on this new company. And depending on the traction that we can get in the next six months, then we’ll probably raise it a round then.
But it’s interesting, because transitioning from a wonderful company like HireVue and working on more of a full-time data science startup, there’s definitely some major changes that happen. One of those changes are, when I’m working on Ziff stuff, being your own boss can be challenging. The challenge there is — a company like HireVue, they’re big enough and they’re a solid enough company that they can afford a few mistakes and a few pitfalls. And for people doing machine learning application development, you’ll definitely run into a lot of those. You might go down a tangent that ends up being a dead end. And when you’re doing a startup and burning resources, you really can’t afford to work on something that’s not critical to work on.
So, a lot of times in data science we like to work on fun projects, exciting projects. You might have ten projects on your list and one of them is actually really enjoyable. And at HireVue we have those options. We have some really fun projects. We did one where we predicted coding ability. As you wrote the code, we predicted whether or not you’d pass the coding assessment. And we showed that we were able to predict that you were going to pass the coding assessment before you’d even completed a third of your coding. So we can already tell you’re a really good programmer once we see a few lines of code. We don’t need you to finish the whole assessment. Projects like that are so much fun, and whether or not they end up in the product is not the main point. It’s the fact that you did it and it was fun and you were kind of exploring. Managing a startup is a lot more stressful because every day is critical and you can’t work on something that’s not absolutely tied to revenue in the near future. It’s definitely a learning process, but we’ll see where it goes during the next year.
Kirill: Yeah. Probably some more grey hairs coming out from your beard.
Ben: Yeah, but hopefully I’ll be the one causing those rather than getting F-bombed by some Manhattan-style hedge fund manager.
Kirill: Gotcha. And quite a few of our listeners actually have ideas about startups in the space of data science and about creating their own business or product or service in this space, in machine learning, data science, artificial intelligence. What would your one biggest piece of advice be for these people?
Ben: The biggest piece of advice would be that it’s a really hot space right now, so for you to go and get funding, a lot of VCs are actively thinking of partitioning funds towards machine learning and AI. So you kind of have a green light there to get in and pitch this, but the thing that we were surprised about is the VCs actually know a lot less than you hope they do when it comes to AI.
One example, during one of our pitches we were referencing IBM Watson and some of these other offerings, and the VCs that we were pitching to, they had no idea what IBM Watson actually did. I think most of the public is probably that way, but I think there’s an assumption if you’re a VC and you’re investing in these tech companies, that you must know in detail what these different offerings do and they don’t. So, I think for people that are interested in pitching their own startup, it’s easier if you find a niche topic. So if you’re going to do like pattern prediction or lead prediction, if you find a very niche topic and you can build a platform around that, something that’s tangible that the VCs can see, that is better.
And then the more you can dumb it down to something—because really, some of the VCs that you’re pitching to, they’ve only invested in like SaaS startups. You might be their first AI startup, so you can’t expect them to know what deep learning is or why your solution is useful. It doesn’t matter if it’s your own startup, but I would go and work for a startup. So if it’s your own, great, but if not, go find a local startup. You’ll learn so much more because you have to wear a lot of different hats and you get a lot of responsibilities and you do a lot of jobs. You might be on more customer calls than you would be, so that was a great thing for me with HireVue. I was able to be on a lot of customer calls and interact on the business side. Where some data science positions you may not get that. So I’m a huge fan of encouraging people to go work for a startup rather than going for maximum job security with a bigger company.
Kirill: Great. Great advice! Thank you so much. And while we’re on the topic of Ziff, would you mind giving us a quick crash course into deep learning? I noticed you changed your LinkedIn profile description as Ben Taylor [DeepLearning]. It obviously shows that you have a huge focus in that area. And deep learning is indeed a very growing part of machine learning. Some companies just prefer to do deep learning than any other aspect of machine learning because they believe that that is more efficient and it can yield even better results. So, if you don’t mind, could you give us a quick introduction to deep learning the way you see it, please?
Ben: Yeah. So, this is definitely very confusing to people who aren’t in our space. So the way I explain it to executives and managers who are not data scientists, because they get confused, is deep learning machine learning is different than what’s going on. So the thing that I tell them is data scientists have a really fancy tool chest in a garage. We have drills, we have hammers, we have all these interesting things that allow us to build some really cool workflow models that provide a lot of value.
Deep learning is a brand new tool that was introduced. It didn’t really become used in the market until around 2014 and after that. A lot of companies don’t have this tool. They don’t know how to use it and they don’t have access to it. But for the ones that do, this tool is so good it is able to beat all classical approaches before it. Those classical examples will fall into a lot of unstructured data examples, so text, image classification, audio classification, video classification, any type of sequential sequence, deep learning will typically do much better than a classical approach like a Bayesian model, bag-of-words.
I can give you some concrete numbers on that. A very popular dataset within machine learning is this MNIST handwritten digit set. So what it is is it’s numbers from 0-9, it’s 28 by 28 pixels, it’s a black and white image. You can literally look at these in Excel, if you wanted to. You could treat it like a table. I’ve got 28 columns and 28 rows, and where the number is sloppily drawn, there are some numbers that are larger than zero. And that dataset has 60,000 images and you can build a predictive model that when you see a new image you will predict what it is – is this a 2, is it 7?
A lot of classical algorithms do this really well. So you can use a Bayesian method, you can use the Random Forest, you could use [indecipherable 32:13] regression, logistic regression, support vector machine. You name your favourite classical supervised machine learning algorithm, you can build a predictive model and you’ll be surprised. It’ll get higher than 80% accuracy and if you’re predicting 10 classes, that’s pretty good since random would be a 10% accuracy.
So they’ll do well but the best you could do—let’s say you’re a genius when it comes to these algorithms. You know them really well and you do all this optimization. The best you could probably do would be 95-96% accuracy. You won’t be able to go above that. Like, I’d be very impressed if you could go above that with classical methods. With the simplest deep learning model, you can get above 98% accuracy and for people that know deep learning well, you can get up to 99.7% accuracy.
And another key difference with deep learning is deep learning does something that is very significant. It does its own feature engineering, where with HireVue when we’re building these predictive models, there’s a lot of human expertise that will go into that. We think this is important in an interview, or we think this is important in a wafer when it comes to predicting yield. Deep learning takes care of all the feature engineering.
The other thing I like to tell people is deep learning does what you do every day. So you walk downstairs into your kitchen and you open up your utensil drawer and you’re going to grab a fork, how on earth did you grab a fork? That’s pretty complicated. How did you find the fork? How did you differentiate the fork from the spoon? And if you look at what your brain is doing, it’s building what I call a feature ensemble. You’re looking at the ends of the fork, you’re looking at the handle. So you have some micro features and macro features and you kind of have this sloppy soup of features that you’ve used to decide that that’s a fork and you’re going to grab it. Deep learning does the same thing. It does a hierarchical feature approach where it can have micro features and macro features but instead of discovering nine features that you found for the fork, it might discover a thousand. Or something much more significant that a human can’t compete with. It’s a longer answer, but I’m so excited about machine learning, about deep learning, I feel like the applications are endless. Deep learning is going to be a major contributor. And that’s a big disappointment for me, when I engage with new students and people looking to get hired and they’re not familiar with deep learning. So I think that’s definitely a requirement.
Kirill: That’s really good. Thank you for that very intuitive explanation of deep learning. It really backs the thinking that deep learning can substitute any other type of machine learning, and it probably will in the years to come. Could you comment a little bit on the most well-known types of deep learning, specifically ANN, CNN and RNN? What do they mean and what are the differences between them?
Ben: Yeah, even though deep learning is surprisingly simple when people understand convolutional networks, there are lots of different types of deep networks that are used. The most basic one that I would use to introduce people to would be CNN. It stands for convolutional neural net. We used convolutions at the hedge fund for performance, but before that I really wasn’t that familiar with convolution. And a lot of data scientists who were just kind of exiting their programs, they’re not familiar with it unless they studied deep learning.
So a nice way to introduce the convolution would be to think of it as filter. So if I have a one-dimensional stock signal and it’s very, very messy—maybe you’re trading Forex so it’s Euro/USD. It’s going up, it’s going down, it’s going up, it’s going down, it’s all over the place. A very popular trading thing to do with that signal is to smooth it.
And to smooth it I would do a simple moving average where I’d have a window of ten numbers or a hundred numbers where I’m sliding it along and I’m doing this moving average. It’s interesting, the fastest possible way that I could do that moving average is not using matrix math. It’s actually a convolution. So, if I do a convolution, that is the fastest way that I can do that moving average. But it can also do more things than a moving average.
It can find like a head and shoulders signal where it’s oscillating up and up and I get a shoulder and I get a head.
And then the second shoulder is just below the head. That’s a classical signal for a pivot point where you might want to short the stock or sell it and that’s a really complicated signal to have. So if you went to some new students and you said, “Hey, I need you to write a computer program that will detect a head and shoulders signal for me,” the problem is sometimes the head can be skinny and sometimes the shoulders can be large. Sometimes there can be spaces and sometimes there can be like a second shoulder that’s not really a shoulder and there’s noise. So that would be a project for them to design an algorithm that would tell me if there’s a head and shoulders signal.
But with convolutions you can actually just solve for the numbers in the convolution and it will detect that. And convolutions on their own, they’re a pretty complicated topic to just discuss through audio without some type of visualization. There are some really good YouTube tutorials. I’d look up “convolutional nets”. The other way I like to think about a convolutional net is think of it as a river. So, you’re up in the mountains and you find this nice laminar flow stream, it’s very calm, you can talk, you know, it’s not making a lot of noise. That is the top of a convolutional net. So I’ll throw an image into that stream. I can explain what’s happening to the image, it’s not very complicated. But as that image moves down through the convolutional net, it becomes extremely complicated where I have filters that are literally creating new images and then those images are becoming multidimensional and then they’re reacting on themselves downstream and then I have different feature sizes where at the bottom of the net it’s like a raging rapid. It’s very complex. There’s a lot going on and there’s a lot I can do with it.
So, you mentioned the other ones. You have LSTM – long-short-term memory. So, if you have a sequence that you’re trying to remember, and a lot of times that can be language, there are some really fun examples where they’ll actually have the computer auto-complete Shakespeare or maybe auto-complete a meme or something. And then you have other complicated nets for speech. There’s one called bidirectional recurrent neural nets where—speech is complicated, because if I say a phrase and you need to understand that I said, “The cat is fat,” so if I say that but you’re not sure that I said, “The cat is fat” or “The hat is fat,” you’re going to decide that the cat is fat because “The hat is fat” doesn’t make sense. And the reason that doesn’t make sense is because you’re actually using a very impressive language model to decide that a hat can’t be fat. So therefore I must have said that the cat is fat.
And to do that is complicated because not only am I using words in the past to make that decision, I’m also using words in the future. Based on how the sentence completed, I’m using the word “fat” in the future, and I’m using all of that to make a decision on what that word is that was unclear, it was hard for me to hear. Bidirectional recurrent neural nets do the same thing. They use information in the past and information in the future to decide what letter was spoken at that moment in time and those are some of my favourite nets right now, are the CTC bidirectional recurrent neural nets. I think Baidu gets a lot of credit for bringing their attention out. A lot of times with recurrent neural nets, you’re using them for some type of temporal sequence, some ordering. With convolutional neural nets you’re using them for spatial information on an image and you can even use them for audio. So, for audio, sometimes we’ll do like a spectrogram, so if I want to detect if you said a single word—like, think of a command. I want to talk to my house and I want to say, “Lights off” or “Lights on,” it’s a single command. I could do that with a convolutional net on a spectrogram. So the audio clip would go through a spectrogram and become an image. I apologize if—
Kirill: No, no, that was great.
Ben: If this is confusing, please take a look at some of the convolutional net or deep learning introductions. I even have a video where it goes through some examples in code. I can shoot you a link.
Kirill: Yes, please. We’ll add it to the show notes. That’s awesome. And that example, I love your intuitive explanations. They really help kind of understand, not necessarily in-depth how the algorithm works, because of course they’re much more complex than that, but just to understand in which direction people should start thinking to better grasp these things. And the example about “The cat is fat,” sometimes in life we have these situations where somebody is saying something and you’re having a conversation and you nod along and you kind of pretend you understood something, but it actually takes you 5 or 10 more seconds for your brain to process what they said because you didn’t hear it that well, and then based on what they said further down, or even just if you give your brain some time, you’re like, “Oh, that’s what they said!” Your example reminded me of that. I wanted to ask you this question: Is it true—because you’ve done chemistry and you’re in artificial intelligence and machine learning now—is it true that electric circuits are one million times faster than biochemical ones so that machines, given the same level of intellect, would think a million times faster than a human?
Ben: I think we’ll get to that point. The other thing that you’re dealing with is bandwidth. So, if you think about some of these GPUs, they are capable of 11 teraflops in a single GPU now, and there’s a really cool graphic that shows when are we going to surpass a human’s ability. So this is interesting. When it just comes to how many operations per second can you do, a computer will definitely surpass a human in the short term. So, we’ll have neural nets in the next 3-4 years that will have more neural connections than our own brains. There are some tasks that computers can do exceptionally well. With image classification, the humans have already lost.
And I’ve had some flak for this, but I’m a huge believer in some well-known respected disciplines and jobs just going away in the future. I feel like the radiologist and the pathologist who in the past have been really respected—you know, my kid has a tumour and we go in and we look at a CT scan or an MRI and this wonderful radiologist with so much experience is looking at this and they’re making a decision. I’m to the point now that in the near future, I’d actually be upset about that. I would be angry why do I not have deep learning that is comprehending a million images to decide what disorder I have. There’s a lot of people that die on a regular basis because the physicians are unable to diagnose these really rare disorders because they personally have not seen it, but a collection of physicians have seen it. So machine learning will be great at that.
Kirill: That also reminds me of—they just released an app recently which you can download on your iPhone or Android and just take a picture of your skin where you think this looks weird and it’ll tell you if it’s a cancer cell or not, if it’s a cancer spot on your skin or not, and they actually tested it against 18 other doctors who look at your skin, dermatologists, and it did as well as they did. So you can already get that on your iPhones and Android devices.
Ben: Yeah. I love that. I think that’s great. So there are some things that computers can do really well. The thing that computers still struggle with are some of the long-term strategy planning. The ones that I’m really excited about right now—we’re going to do a demo this year with some computer games for Ziff. Deep nets can play games like “Doom”, these first person shooter games where they’re walking around, and the deep net, they’re controlling in real time what the player should be doing as far as navigation and shooting and things like that. And the deep net can do it better than a human. And the other example we’re seeing is with some of these drones flying. A drone flying through a mountain trail—they’ve already shown that these deep nets can predict where to go. They’re able to do it because they’re training on a short-term reward. So if I’m playing a game and I shoot someone in the game I get some short-term reward. Or if I’m flying a drone and I crash or I don’t crash, the short-term reward is very clear.
But if I’m playing a strategy game where I don’t get a reward until the end of the game, or I’ve gone through some fantasy level or something—“Starcraft” is the one that DeepMind is focusing on right now. Those are very hard. They’re very challenging. So the idea of drones flying around for the military and doing short-term actions, whether they’re surveiling a region or they’re taking out someone carrying an AK-47 who’s not allowed to be carrying an AK-47, those types of actions are very simple. Computers will do those exceptionally well at faster frame rates than humans ever could. But when it comes to making long-term strategy planning decisions that’s still quite hard for computers.
Kirill: But we’ll get there eventually, right?
Ben: Yeah, we’ll get to Skynet soon enough.
Kirill: (Laughs) And that’s my next question. You’re very knowledgeable on the topic of deep learning and artificial intelligence in general. There’s this futurist in the U.S. I’m sure you’ve heard of, Ray Kurzweil. And he’s got all these predictions about when and how and where we will get to with machine learning and artificial intelligence. And he’s predicting that by 2030 we will have amazing things. Like, we’re going to have nanorobots in our blood stream looking after us, making sure we’re not sick, we’re going to have self-driving cars and everything is going to be connected, the car is going to redirect you to the hospital at the first sign of a possible heart attack, and so on. But also he predicts that somewhere around 2050 and onwards, AI is going to be so dominant and so strong that it’s going to start making decisions for itself, it’s going to start thinking of where the world should be going and it’s going to start disregarding us or even eliminating us. What do you think of this whole AI revolution topic? What are your views on this subject?
Ben: Yeah, I think that people that know me well tease me that I’m already an insider with the robot empire just because I’m such an advocate for accelerating. I don’t think humans should be taking the gas off. I think they should be dumping the gasoline on the fire and make this reality happen sooner. This is getting up to 2040, to that type of ideal scenario. And the interesting thing about this is this will happen whether people want it or not, and the reason is because of the good that people will see, the good in the security that they’ll see upfront.
An example will be driverless cars. You know, so many people die because of people driving. You have drunk drivers, you have people that fall asleep, people texting on their phones. In our own cities and counties, we have fatalities on a regular basis from human mistakes from driving. So, self-driving cars, that seems like a clear win. In medicine there are so many applications to save lives beyond what medical researchers have been capable of in the past. The one that I like, in the U.S. there’s been a lot of heated debate between this “Black Lives Matter” movement, so “Black Lives Matter” versus “White Lives Matter”. Being on the machine learning side, the way I see it is people are missing the point and the point is that humans will always be bad when it comes to making split second decisions that affect their lives. So, if you’re a police officer and you’re called at 2:00 in the morning and someone says, “There’s a kid in this park that has a gun,” and getting back to that limited experience, based on that you had a very negative experience with this particular minority, and you go to the park at night and you’re on the edge, you could be shot like your partner was 2 years ago, and this kid has an airsoft gun or they have a toy or they don’t even have a gun and they’re just walking around in a hoodie and you shoot them—these things happened way too many times.
So you have this “Black Lives Matter” movement, you have the “White Lives Matter” counter-movement or “All Lives Matter” counter-movement, and what I think is people need to realize that humans will struggle with this, there’s not enough training. There’s no amount of training that will fix this. But the thing that will fix this is—imagine this scenario: You get the call at night, the police officer shows up. He doesn’t even get out of his car and a small drone detaches off of his car, flies over this individual and does a threat assessment. The technology exists now, using deep learning, where you could do a threat assessment better than they could have in the middle of the day if they were not stressed and they took a few minutes. So the drone will do a better threat assessment. The police officer doesn’t care. They’re eating dinner in their car and the drone comes back and tells them that there’s no threat and then they leave.
So the police will appreciate it. For the minorities that have been negatively impacted, they will appreciate it. So the idea of AI security for the police, I see that as being a huge benefit, but I think that begins to scare people because if there’s a mass shooter—like, you know, in your country and my country we’ve had mass shooters—if there’s a rogue human there, I can promise you that in the future we will have humans that decide they want to kill a lot of people and we can’t prevent that from happening. It’s kind of that outlier thing. Like, most of the humans in the world are great. We love them, they’re fantastic, but you’re going to have an outlier and how much damage can that outlier do.
But with machine learning, if you have a drone that determines that someone is shooting someone on a campus, you probably want that drone to do something about it. You probably want them to incapacitate the person. You know, you’re going to start to cross all of these ethical lines of “What do we feel comfortable with?” You know, you could easily have a drone where, as soon as someone fires a shot, within two minutes they’re incapacitated, whether it’s lethal or non-lethal. So do we feel okay with that? I think it’s fascinating. I’m just going to wait and see. Let’s see what comes in the future, but I’m a big advocate on saving lives and improving. But I think you will see riots in the future against AI around job loss. In the U.S. we’ll have over 3 million professional drivers that are out of work and unqualified for other jobs because of AI, and you’ll have lots of other industries that suffer as well.
Kirill: Yeah. It’s very interesting how Uber introduced all these cars and now has all these drivers and within a few years they’re going to replace them with driverless cars. So they created an industry, created a lot of jobs, and now it’s all going to get replaced. I agree that there will be a lot of benefits coming from artificial intelligence. But I was watching a TED talk recently, it was by Sam Harris, “Can we build AI without losing control over it”. Have you seen that one?
Ben: I haven’t seen that one, but it sounds fascinating.
Kirill: It’s pretty cool. The way he describes it is—so, we’re creating AI and we want to improve our lives and so on, but as an inevitable by-product, we’re going to create this super intelligence that’s going to be smarter than us and it’s going to think way quicker than us and so on. We don’t know what it’s going to want to do with us on this planet and how it’s going to react to humans being here. It’s going to come here around 50 years from now.
One really interesting comparison he says is, a lot of people understand this, a lot of people appreciate the concept and we’re just kind of waiting to see what happens when this AI does come. And he says, what if we got a signal from an extra-terrestrial species that simply said, “Humans, get prepared. We’re coming in 50 years’ time?” Obviously we’d start doing some preparations, we’d build defences or come up with some sort of strategies on how to deal with aliens and so on. Here we are creating this alien ourselves. It’s coming. We know, based on Moore’s Law and other predictive mechanisms, we know that it’s coming in about 50 years from now, but nobody is doing anything about it. Does that concern you at all?
Ben: No, not really, because even for the most advanced AI that we have now, and even for the future AIs, like the AI 50 years from now, there is always a concept of an objective or a goal. There is always a mission and the mission is defined by the humans. So the joke on the AI side is the mission will be human safety and prosperity. Like, that is the goal of AI and that is hard-coded into the kernel. A lot of these AI movies, they talk about scope creep where the kernel has been rewritten and it’s able to jump out of that objective. So the computer is really a slave to whatever the human wants as long as the objective is well-defined.
And I’m probably embarrassed to admit this, but I’ll admit it anyway—one of my favourite quotes or taglines that I’ve had in the past was, “the proudest moment of my career will be when the AI kills me because of an objective that I’ve overlooked.” So the joke around there is, I love AI, I want to design it but—I think in “I, Robot” the robots eventually kill the scientist and when they kill the scientist, what they’ve done is—the scientist is the AI expert and he’s defined the objective and the rules but the AI has figured out a way to satisfy his objective. Really, the AI is thinking of something that you didn’t think of, so if you kind of extend that joke, it would be that if I tell the AI that all I want is for my wife to be happy and then the AI realizes that in order for that to happen it must kill me, it has outsmarted me. It’s something that I never saw coming because I felt like that was not a possibility.
We already see this with computer games where the AI will begin to cheat. So if you have AI playing a computer game, if there is a way for the AI to cheat, and they’re seeing this with Mario and a few different things where the AI is able to jump on the corners of boxes and things that a human would never do. But the AI has figured out that it can do it and it gets an advantage. So just like that, the AI has this insatiable appetite to satisfy this objective. It’s really the fault of AI. So AI will cheat if it means they can satisfy this objective better.
Kirill: Yeah, because it doesn’t really care about the rules unless you tell it that it can’t cheat.
Ben: Yeah, you have to define all the rules. So, if drones need to get a faster response time to a gunshot victim and the drone realizes it can fly through a window and break the window, it will do that, but you might say, “Wait, hold on. I didn’t want you to break through all the windows on campus to go get this perp.” Like, I do but I don’t. You really have to define the rules well or it will surprise you.
Kirill: Okay, gotcha. Wow, that was a crazy discussion. I really enjoyed that. Thank you very much for coming on the show. This brings us to the end. I would love to continue going but maybe we can have you on some other time when your startup is full throttle and you can share some more insights. How can our listeners follow you and contact you and get to know you better if they would like to know more about your career?
Ben: Yeah, so LinkedIn is probably the best, so “Ben Taylor Data” and they can find me on LinkedIn. I am on Twitter. I’m not very good at Twitter so if they do send me a message on Twitter, maybe in a few months I’ll get around to responding. Yeah, so they reach out to me on LinkedIn. I do get a lot of requests for mentoring and for job help or getting a job, and so sometimes it’s hard to respond to all of those. I’m more likely to respond if I know that they’re not asking a question that’s already mentioned. For my blog post, if you read my blog post and there’s something that’s not addressed, please let me know, whoever you are, and send me a message and I’ll try to respond to you. If there’s something that’s not addressed in my blog post I’d be happy to write it up or discuss it with them. But if you’re asking me, “How do I get a job as a data scientist?” and that is something that’s heavily discussed in all of my blog posts and I’m less likely to respond. Yeah, I try to respond to people when I can.
Kirill: Yeah. And a quick heads up for our listeners, Ben won’t admit it but most likely it is an AI responding to you anyway, so don’t keep your hopes up. (Laughs) Seriously, in terms of articles, I highly recommend the “Data Scientist Type-E” article by Ben. You can find it on LinkedIn or just search, google “Data Scientist Type-E” and we’ll also include it in the show notes. It’s a fantastic article. It completely changed my career and my perspective on data science in my life. So thank you so much, Ben, for that. I became a huge fan right away after reading that article. Another good one is, “Why Working Hard…”—what’s the correct name there?
Ben: “Why Working Hard Won’t Make You a Data Scientist.”
Kirill: Yeah, great article as well. I really liked how you described a good tip for people, a silver bullet for people to become great data scientists is like, get your favourite Python libraries or R libraries and dig into them, find all the code and make corrections and submit those corrections. One of my friends actually did the same for Linux. He found a bug in Linux and submitted the correction, they accepted it and he put it on his resume and got a job with Microsoft.
Ben: Oh, yeah, that stuff can be huge. I did have some pushback from a researcher in Switzerland that was pretty frustrated that I encourage people to do that. The reason he was frustrated is because, yeah, if you’re digging into code for an FFT or support vector machine, like the real guts of it, you’re not going to get any value out of that. You’re definitely not going to do anything to improve it. But I’m thinking more about the high level stuff like pipelines and Pandas, like Pandas Get Dummies. Dig into that stuff and understand it at a high level but I don’t expect you to be digging through an FFT details. There’s definitely value there, but don’t get bogged down on anything that you think is too complicated. Just the high level will definitely offer a lot of value. Yeah, thank you so much for thinking of me and inviting me on the podcast. Hopefully it’s useful.
Kirill: Yeah, definitely useful. It’s a great pleasure to have you. And one final question. What is your one favourite book which you can recommend to our listeners to help them become better at data science?
Ben: This will be an unusual book recommendation, I think, because it’s not a data science book. But one of my favourite books is “The Lean Startup” by Eric Ries. The reason I love it, and I think a lot of data scientists can benefit from it, is it is so important for us to fail quickly and a lot of us struggle with that. So, a lot of data scientists, they’ll get a problem and they’ll immediately start spinning up something that is much more complicated than they should. Really, the thing you need to decide is, “What’s the fastest way for me to fail at this particular problem?” And the reason that’s a benefit is because it’s also the fastest way for you to decide if it’s worth your time.
So, if you’re spending a week or two to decide if this sophisticated approach is important, then I see that as being a waste of your time. But if you can decide in an hour that this is useful, then by all means do that. The other thing I’ll tell people is, even though you have much more sophisticated tools and methods out there, you can get 80% of your value with the Bayesian method or logistic regression, so please do that. You can do that very, very quickly. And if you’re dealing with a custom dataset that’s very messy and it requires a lot of custom munging, just run it through Pandas Get Dummies, which is a very quick auto-tokenizer or munger, just to see if there’s value there. Because if you can see that there’s any value, that can motivate you to do something more complicated. But definitely don’t start with deep learning LSTMs. I’m a huge fan of Eric Ries’ idea that you fake an MVP or just do everything you can to fail quickly or get customer feedback. This has happened at HireVue. We’ve built some very complex things using Siamese nets, like top-notch technology, and then it turns out that it was more of a curiosity for the customer but there weren’t real dollars tied to it, and so that thing kind of died on the vine. Thinking of what the customer wants can be different than what the data scientist wants. So the better that users can think of what the customers want, the better off they’ll be at whether it’s their own job or their own startup.
Kirill: Gotcha. Thank you so much. “The Lean Startup” by Eric Ries. It sounds like a great book for data scientists to save some time and really understand what the best approach is very quickly. So thank you again so much. I really appreciate you coming on the show. I’m sure so many people will pick up so much value from here and hopefully they will learn even more from your articles and the future ones that you’re going to write as well. Thank you so much for everything you’re doing.
Ben: Yeah. Thank you. Thanks for having me.
Kirill: So there we go. That was Ben Taylor, Chief Data Scientist at HireVue. I hope you enjoyed this podcast. As you can tell, Ben has so much knowledge and value that he can share. This podcast could go on for hours and hours and hours, but hopefully we were able to convey some of those very powerful insights that will help you understand a bit more about deep learning and what it actually is and how it can fit into your career and into your startup or into your organization and into your role that you’re performing right now.
And also, Ben shared some great tips all around, including on how to build up your career. My favourite part was when Ben described and drew those numbers on how deep learning is so superior to other machine learning algorithms. For instance, when you take the MNIST dataset with the hand-drawn digits, with a neural machine learning algorithm, any one of your choices, you can probably get an accuracy up to 95%, or if you’re very good and you try really hard, up to 98%, but with a simple deep learning algorithm you can get an accuracy of 99.7%. That’s just crazy. I think that’s a really good example of how deep learning is superior and it’s something for you guys to think about.
So that brings us to the end of today’s podcast. Please have a look at Ben’s articles on LinkedIn. I highly recommend them. They can really change your perception of your own career and how further you can develop your skills in the space of data science. You can also get the show notes for today’s episode at www.superdatascience.com/29. There you’ll get the transcript, all of the resources and links mentioned in today’s session, and also a URL to follow Ben on LinkedIn or Twitter. And finally, if you enjoyed today’s conversation and you’d like more episodes like this, then you can subscribe to the show on iTunes, it’s the Super Data Science podcast, and that way you will get the freshest and newest episodes as they’re released. And I look forward to seeing you next time. Until then, happy analyzing.