What Fascinates Ameen Kazerouni, Data Scientist at Zappos?

Bobby: This episode is yet another glowing example of how unpredictable and interesting one's career journey can be. Stick around until the end and find out the first thing amine ever bought on Zappos. This is the very first time we are doing the Loka podcast remotely.

Usually I have the good pleasure of being face-to-face with my guests, but in this particular instance, we decided to try something different, push the edges of technology and do this remotely. So while I'm in my office here in San Francisco, Ameen is over at Zappos, I believe in Nevada. Is that right?

Ameen: Absolutely. Las Vegas.

Bobby: Alright. Las Vegas, Nevada. So without further ado, why don't I let my guest fully introduce himself.

Ameen: Hi, my name's Ameen Kuni. I currently lead the data Science and Data Science platform teams at Zappos. First of all, thanks Bobby for having me. This is very exciting. I’ve enjoyed a few of the other episodes and I'm excited to be here and talk through data science.

A little bit about myself before we jump into the nitty of AI and machine learning. I started off doing my undergrad in just regular computer science, which is where I was first exposed at the University of West Georgia. It's a tiny university in Carrollton, Georgia. It's got a fantastic CS program. And what was really exciting about it is the chair of the department at the time had this vague interest in robotics and genetic algorithms, which surprisingly the genetic algorithms bit came back much later in the career, in my career, and I'm sure we'll get to that.

But I did a few independent studies with him where we built fun robotic projects and. That's where my first brush with machine learning and AI happened as an undergrad. Over there we published some papers on using robots as tools to teach search algorithms and teaching robots how to walk like human beings using the Darwin anthropomorphic robot that they had back.

I don't know whether that's still at play in the robotic community, but that's when I got first exposed to int intelligence systems and it was very exciting to me. But at the same time, both my parents are doctors, so there was still this little piece in of me that was like, I too will be a doctor, not a computer scientist.

That's what led me to get into this hybrid field in grad school at Emory where I was studying biomedical informatics, it was pretty much a kind of detailed dive into using ML and good data practices to access data from various modalities like flow cytometry data, imaging data from MRIs, genomic data pathology reports, which is just natural language from physicians.

It's a hard charge of every different type of data you can get access to. And then you gotta like figure out how to use it all together. So we did a lot of cool projects there. I was pursuing my PhD at the time. It was pretty intense. It was a good solid three years and it laid down the foundations for most of my academic ML background.

And Dr. Ashish Sharma there was a fantastic guide to me and we worked together on a variety of different projects, which we can jump into a little more detail as we continue chatting, but it laid down the framework for being able to use different ML techniques for different types of data, and understanding how much diversity there was in machine learning as a field independent of the domain you're working in.

And I think that's what kind of sparked my interest too. Go out and do an internship at Zappos and get a feel for how could I use these same techniques we using here in a completely different domain. That's what got me super excited. I saw an opportunity to get into this e-commerce space and see how we can use the same techniques that I've been practicing over the years in robotics and healthcare and see if I can port very similar methodologies to e-commerce and leverage ML cross domain without having to change the techniques.

And that's what led me to Bumblebee Analytics, which was a small startup in Bombay, in India, where we dealt predominantly with e-comm clients trying to build big data infrastructures and things like that. And soon after that I came back to Zappos to start their digital experience data science team.

But since then, it's now one of the largest tech teams here. It's doing some pretty cool work. We've got some very exciting projects currently powering the customer experience using AI and ML on the website, but it's been quite the ride so far. And I guess that's the introduction. That's me in a nutshell. And that's the various areas of ML and domains in which I've slid about in this field.

Bobby: Super interesting and near and dear to my heart, because I'm spending more and more of my time and thus Loka’s time working on more projects in the life sciences space is something that's near and dear to my heart, in fact.

You should check out the previous episode of my podcast or more on that, which is pretty interesting. So what I was gonna ask was, so when you went to Emory, you were initially thinking about pursuing a PhD then, is that right?

Ameen: Yes. I was a proud PhD dropout there.

Bobby: And you're in fantastic company. It seems to be that's a requirement for successful people is having wherewithal to know when to change direction. You were three years into it. Yeah. So what made you decide to try a different direction at that point and not finish the PhD?

Ameen: So I think for me it was the base of it was something about the biomedical informatics. And this is not anything against research and academia, especially in the medical field. Cause I think it's critical to be able to use things like ML and AI to drive healthcare forward. And I think it's going to be that differentiating factor that. Analogous, takes us to the moon and healthcare effectively.

But for me in particular, I liked the idea of seeing what I was doing immediately impact the domain that I was working in. And I think e-commerce, and that's very selfish of me honestly, but in e-commerce it gave me the ability to develop an algorithm and immediately see its impact on a customer base, like immediately affect the people I was trying to affect using machine learning and mentally. I think the shift from healthcare to shoes was a little interesting for me to make that transition from healthcare to e-commerce in general. But I think it was the pace of it academia, especially in the medical field, requires discipline in terms of being patient, FDA approvals, clinical trials, long drawn-out processes to see these things actually do what you intended them to do. While in e-commerce I've had the opportunity to fail fast and, fail hard, but also have these successes immediately. Give me the gratification I was looking for while practicing machine learning.

Bobby: That makes a lot of sense. Clearly the whole notion of failing and failing fast and healthcare has very different repercussions. But if you're an action oriented person that wants to constantly see improvement in results and get feedback so you get better, those feedback cycles are much more tenuous and longer.

Taking what you learned in this bioinformatics space and using it in different domains, let's stay in the bioinformatics for a second and just for folks that don’t know, maybe just take a couple of examples of the more common types of machine learning algorithms that are used, today in the bioinformatics life sciences space.

Ameen: This takes me back to my first intro to biomedical informatics class that I took back in the day where the professor that ended up becoming my advisor, Dr. Sharma, he gave me this example, which really captured, was very captivating to me in terms of what biomedical informatics was.

And he was giving me one of the first examples of biomedical informatics. And I forget the exact name of the town and hopefully I'm not getting the disease wrong, but there was a town in, I think England, where there was a cholera outbreak and nobody could figure out what was causing it. And what was amazing is that there was, I forget whether it was a physician or not, but the point was that somebody took a map of all the wells in the town and overlaid it on top of a map of where the cholera outbreak was happening.

And was just like, this was an example of medical informatics being used for the first time to identify that these were contaminated wells in the city that were causing the disease. That example is what kind of laid the groundwork for what my research and stuff in the biomedical informatics space got informed by was there are all these different modalities of data that are important.

By themselves, very important by themselves, but an absolute goldmine when you're able to look at them together and being able to view this cohort discovery as we called it, like being able to just query different modalities of data together, gave you a view into problems that you didn't know could be solved because you thought you didn't have the information.

I think that was fantastic for me in terms of machine learning algorithms in particular. In the life sciences space. It's so diverse, right? Just step stepping away from work that I was doing in particular, named entity recognition is a field of national language processing that uses an assortment of different techniques like conditional random fields or even LS dms and recurrent neural nets.

Now, to do. Named disease entity recognition in like pathology documents. And in addition to that, you can like completely switch gears and go to like image segmentation algorithms for identification of cancerous nuclei and whole slide images. And you can go from there and switch gears again and actually look at using much more traditional like classification techniques like random forests.

To be able to make predictions on simple clinical data that you have available to you in just electronic medical records. So what was cool about it though is that almost every single one of those techniques works when you're trying to sell shoes as well. And that's what's really interesting to me about machine learning is how domain expertise really informs the algorithm you choose and how you use it.

But at the same time, you can take what you understood about one domain. And leverage it in another very, got some cool examples for that we can jump into? Yeah, why don't we, why don't we go ahead and do that? There's this cool analysis type called survival analysis. That they used to, I think it was called the West Nile Disease.

I think that's what the name of the disease is. But effectively what the West Nile virus and what they did there was they used this survival analysis to show that people who were infected with it had a lower, a significantly lower chance of survival. Effectively showing. And the reason survival analysis is different from regular, just comparison of two cohorts is that you actually have, let's say during a clinical trial, people drop outta the clinical trial or they don't.

Do the things you expect them to do in the clinical trial. It helps account for those kind of anomalies where let's say if somebody disappears from a clinical trial, it doesn't necessarily mean they passed away. They could have just chosen to stop coming to the clinical trial. So how do you account for that?

So we were actually using machine learning to be able to help predict a customer's shoe size on the website. And that's a feature that's live on the website right now. Where we can guess your size accurately in whatever shoe you're gonna buy. If you've been a customer with us for a while, and the, what we used to evaluate if that was working or not was treating the return of a shoe as the analogy to like the shoe dying.

But someone keeping the shoe did not imply that it fit them. It just imply that it was no longer part of the trial. So it could be that it fit or it could be that they left the trial and we actually used survival analysis, which is. A favorite of the CDC for  survival analysis to figure out the survival rate of shoes to evaluate return rates.

So that was an example of how, and that was it. It gave us insight that we originally didn't have because we weren't just comparing two groups of people return versus not returned. It was like accounting for that. Maybe they just kept it and it didn't fit them. Fascinating.

Bobby: I never in my head would've been able to stitch that together in treating returns in the way that you did there.

That's so fascinating. And I'm guessing there are a lot of examples that are like that where you're thinking about healthcare and treatment and patient outcomes, and you can transform that into e-commerce metaphors.

Ameen: Yeah there's various ways to do that. What I think I found very helpful for me is and this is not maybe like a healthcare technique in particular, but just genetic algorithms are something that we use very heavily.

Bobby: Actually, maybe just for the folks in the audience that don't know, maybe you could just take a step back and explain what genetic algorithms are.

Ameen: Oh, absolutely. Yeah. So a genetic algorithm effectively is a. It's a suite of, it's a, so they're known as evolutionary algorithms. They've got a flavor of randomness to them, and they evolve and adapt to the ecosystem that they're dealing with.

A genetic algorithm in particular is heavily influenced by the basic philosophies of evolutionary biology, which is that there's two principles that are used in the algorithm development. One is that the more fit an animal is, the more likely they are to reproduce. It's kinda like survival of the fittest, which—Darwin did not say that. I think it was an economist, but that’s besides the point. But yeah, the more fit a parent is, the more likely they're to reproduce. And the second principle is that traits of the parent carry forward into their children. So very simply put, those are the two core principles that drive genetic algorithms and the fitness or how fit apparent is in the evolutionary senses, merely survival.

But in terms of an algorithm can be modified to be any kind of objective function. May it be increasing sales, may it be increasing conversion, may it be increasing search page clickthroughs. And you can use that fitness function to search a seemingly impossible to search large dimensional space to identify an optimal solution or a candidate solution that never occurred to you.

So the algorithm evolves adapting to the ecosystem and finds an optimal solution. That may not have occurred to you. You're not saying, oh, here's six options. Let me figure out which one's the best. All you are saying is, here's the problem. Here's three guesses. Find me a thousand solutions and figure out which one's the best.

And it like that evolutionary technique. So it's quite phenomenal to watch evolutionary biology at play. At solving a search algorithm.

Bobby: Okay. So would the, if you were trying to take that a step further and give an example of using this genetic algorithm to solve a real world problem, would a search algorithm be where you would start or is there like a, another

Ameen: one that's, there's a much simpler example.

I think the search algorithm is something that we've actually been working on at Pure, and it's been, it's probably one of the hardest things we've tackled and it's been happy to say it worked, but I think a simpler example may be better for someone who is being the genetic algorithm for the first time.

So I think a lot of the listeners who are interested in AI and ML and just general optimization are familiar with what's known as the traveling salesman or even the nap sat problem, right? Where you've got a bag that can hold 20 pounds and you've got 50 books, and the goal is to get the maximum number of books in this bag.

That, and the books have varying weights. So let's imagine a simple list of zeros and ones of 50 of them. One for each book, right? Every time. One of them's one instead of zero. It means the books in the bag. Now, if you try and think of every possible candidate solution, it's a ginormous number of solutions as to all the combination of books together that are under 20 pounds that can fit in this bag, right?

And you can't possibly go through every solution. But a genetic algorithm will allow you to start with a random combination of books that are in the bag, maybe two or three candidate solutions, and then you start the evolutionary process and it figures out which book being in the bag is most useful for being able to fit more books in the bag.

So let's say there's a 15 pound book, right? Every candidate solution with that book in the bag is probably gonna fail because you can only fit books under five pounds in the bag. Now you've already taken up. 15 pounds with one of the books. So that book as a genetic trait is slowly going to die off right in the evolution of this solution.

All I can find that, oh, I've got 40 of these books in all. One pound each. Those are gonna be really strong genetic traits because you already got 40 books in that bag and you've still got 10 pounds to play with, right? So those genetic traits are gonna keep propagating downward from population to population and probably end up in the fittest surviving solution.

Oh, interesting. And that's how, it's cool to think of it that way. Thinking of each book being in the bag as a genetic trait, like black hair or sure. Baldness and stuff like that. And these traits either carry forward or don't. And you can use this concept to optimize for any fitness function.

Bobby: So the NASDAQ problem is one that is easy to kinda understand cause it's famous. But in terms of and this can be outside of Zappos, but just in e-commerce in general, let's say, are there, more famous examples of how companies have utilized genetic algorithms, have you know, better outcomes for their business?

Ameen: I actually think that. We are one of the few people using it in an at scale solution. I've given a few talks about it and actually done another podcast focused purely in genetic algorithms. And I can tell you why I think they're useful in a situation like e-commerce, right? Any machine learning algorithm that you build in e-commerce has to account for a changing customer base because you've got a mix of loyal customers, new customers, first time purchasers coming into your website.

Then you've got seasonality. It's cold. People want boots. It's summer. People don't want jackets. It's people want flip-flops. And then you've got a constantly changing inventory, like what do you have in stock? What do you not have in stock? So a machine learning algorithm that you train at one point of time works for that state of your ecosystem, of your customer base, your seasonality for the domain you're in, the inventory position that you currently hold, right?

Now summer becomes winter. Your inventory changes. The customers you have either moved on or you've acquired new customers. And the general seasonality for whatever it is you're selling has changed. What do you need to do here? You need to build a brand new machine learning algorithm. You need to retrain it to account for that seasonality.

While if you've got a genetic algorithm that's constantly running and constantly evolving, it's tremendously valuable cuz it adapts to the ecosystem. As evolution allows you to so it adapts to the ecosystem. And a great example there is a search algorithm like being able to surface the right brands and the right products to the right customer.

Given the current seasonality, given the current time of year, given the customer's current preferences and their most recent behavior is a valuable insight that a static machine learning algorithm cannot provide you.

Bobby: Yeah, so it can be, it's much more adaptive and therefore, the value of the query that it's hitting the mark of what the person, what the intent was, is much, much higher.

Cause you have a lot of moving variables. Yeah, that's pretty fascinating. So I wonder, I'm sure, I'm guessing Google probably uses a very end of this for their own queries. And so you could be the same guy and type in the same query, and if you did that 365 days a year, even from the same IP address, it may be a different result based on these other environmental factors.

Ameen: Exactly. It definitely would be a candidate solution for that kind of problem.

Bobby: Fascinating. Okay, so let's zoom out again and use kinda your fantastic experience on this to do another primer, which is, imagine being executive coming into this world and hearing and seeing headlines about artificial intelligence, machine learning, and deep learning. Just those three terms. Okay. AI, ML, and deep learning. How would you just describe at a high level how those things are connected and what the diff how they're related?

Ameen: Oh man, that is one of my favorite questions because it's asked so often, and I'm sure you've asked this question to other people as a, and my guess is you've always gotten different answers.

At least that's what, when I ask people that question, I always get different answers, right? So I'll give you my personal opinion here in terms of what the core differences are. So I think between ML and deep learning it's very simple, right? Deep learning is like a subset of ML. Your logistic regressions, your XG boost, your random forest.

Your clustering is all machine learning, right? So machine learning is broken up in a supervised and unsupervised learning. And now you've got reinforcement learning as well. Supervised learning is effectively when you've got historical data that, and you have tagged historical data. So let's say I'm trying to use all of someone's click-through behavior to predict if they're gonna convert on the website or not.

I've got historical indicators of what people have done on the website and whether they convert it or not. So I can train a model on that and then predict on unseen data points when someone's live on the website. And see if they're gonna convert or not. So that's supervised learning. Our supervised learning is when you're like doing things like clustering.

So algorithms that fall into this family is means clustering, hierarchical clustering. These are techniques that basically allow you to create groups within your data points without actually having a label associated with it. So an example there is, if I told you, oh, I have something that's yellow and long, it's probably a banana.

But if you didn't have the tags and you just had a bunch of fruits and you try to cluster them, you'd probably cluster based off of color, taste, shape, and create like subgroups and you'd be able to like assign hopefully names to them. Or if you're using cars, you could use different principles and properties of these cars and get oh, these are vehicles you could get, these are airplanes, these are boats. These are cars. These are trucks, these are sedans. These are hatchbacks, and this is a fighter jet, a commercial airline, a speedboat, streamliner. And these kind of become like this is known as hierarchical clustering. You keep breaking it down, right? This is when you don't have a label. And reinforcement learning is a little less known of than your classical supervisor and unsupervised learning.

But it's effectively when you've got like semi-supervised data, right? You've got some label data, you've got a human actor informing the algorithm as to whether it's right or wrong, and it just keeps learning from that. So this is the family of machine learning, right? Deep learning or artificial neural networks and the variety of artificial neural networks that are out there.

Generative adversarial networks, recurrent neural networks, long short term memory, LS, DMS, et cetera, et cetera. Are all the hotness in machine learning right now? Those are examples of supervised machine learning techniques, right? They have historical data, things that they've, and they have a label, and you're able to learn from that historical data to be able to make a prediction on something you've not seen before.

These deep learning techniques are, they're more black box. They've got multiple layers in the neural network layered together. It's not as, it may be not as interpretable because you don't know what's going on inside the various layers of the neural network, but it's basically like using all the various combinations of your input data to make a black box kind of prediction.

It's able to learn patterns in relationships that you wouldn't normally be able to do in a kind of a non-linear technique. Now AI is where it gets a little difficult to come up with a definition. I've heard a variety of definitions. One being where AI is any kind of artificial intelligence, right?

There's this fantastic data scientist, his name is Pedro Alves. I worked with him a lot in the past, and he gives an example of AI, which is a paper towel dispenser, right? It's just mimicking intelligence, right? Like when, if I asked you for a paper towel right now with my hand, you'd give one right? Go to a paper towel dispenser in a bathroom and you stick your hand underneath it and it like spits out a paper towel that shows some semblance of intelligence, right?

Technically that's artificial intelligence right there. Or you can see AI as like the general AI where everyone's oh, the singularity, and we are gonna get machines that can act like human beings. So literally completely two opposite ends of the spectrum can be used as a definition of ai. I personally believe that.

The theory behind machine learning and AI came long before the hardware that allowed us to do AI and machine learning. Back when that had happened, they had called it AI and then it fizzled out because it got super exciting and then no one could do anything with it cuz the chips didn't exist to do it.

So what happens when you've got technology that you can't use? You get sci-fi movies and sci-fi movies aren't like, oh, look at this AI in this beautiful world and just help. It's The Terminator, right? So you've gotta buy, you got a lot of bad press associated with ai, and then suddenly we got the chips to be able to actually do the neural networks and stuff. So we just rebranded it as machine learning and did it anywhere.

Bobby: And it became, yeah, and it, and the rest is history as they say. Per your prediction, I think I get asked it and I ask this question quite a bit more and more these days, and I definitely do get different answers, but that's definitely one of the better ones I've heard that resonate with a lot of my own feelings, so That's great.

I appreciate you taking the time to do that. So if we then dive a level deeper and start talking about, just basic algorithms and you. I'm just curious on your thoughts. If you take something like gradient boosting, that's something that's I think has been around since the late nineties, but has been picking up steam much more recently.

Like to what do you attribute that, decades gap in between, discovery and mainstream use?

Ameen: So I think this is, again, answer that I think is a personal opinion and we do a lot of gradient boosting as well. It's basically optimizing a cost, right? Gradient boosting is effectively optimizing some cost and it's like trying to like maximize something, right?

So very simply put like a good example of gradient boosting is like tree boosting, right? You've got a lot of different gradient tree boosting. It's used for like classification, regression, et cetera. The advantage that I've found there is twofold, right? One, it's interpretable. I think as ML has picked up, there's a lot of companies that do a lot of research in AI and ml.

But you'll find that a lot of companies have data and they don't really know what to do with it, right? Like they have a lot of data, but they don't have the requisite talent. And even if they did get the requisite talent, it's very difficult to deploy neural networks and these deep learning techniques in production and environments, right?

So if you look at Zappos as an example, you look at any other retailer as an example, or you just go and look at the field as a whole. And you or you run a Google search for, how do I deploy a long short-term memory network at scale in a Java api? It's really hard to do that. There isn't like stock packages out there, and I give a talk on this on machine learning in production, which talks about a few ways to navigate around this problem, but there's not a lot of well defined methodologies out there, but gradient boosting or the gradient tree boosting provides you with solutions that can be represented in.

Standardized formats like predictive model markup language MML, which allows you to use JPML in a standard spring boot application where API developers are familiar with what they're doing. You don't need machine learning engineers writing your production API code. They're able to take the knowledge of your machine learning engineers and put it in a production setting.

You're no longer looking for those unicorns who are like at scale software engineers and deep learning specialists at the same time, and more importantly, you understand the why. In a gradient boosted algorithm, you're able to actually know what variables are influencing your output and how, and it's not just a black box technique that you need to rely on where there's a lot of research being done in increasing the interpretability of neural networks right now, but it's not quite as simple as, here's a tree: if it's greater than five for this variable, go this way, otherwise go that way. And ease of deployment I think are critical in making them so popular now that people are starting to ask the question, how do I use machine learning in production?

Bobby: The accessibility to it was something that was pretty true in the late ‘90s. But I think the difference, building on what you're saying, it, it seems to me that I think there were these two big step function changes in the power of computing that then made them become popular in the last couple years. Something like grad boosting, like why did gradient boosting not become like a known term in the late nineties versus like now? And I think it's one, the processing power of chips, which you alluded to earlier, was not there.

And the second thing I think is just the explosion of data itself, right? The amount of data that is out there in 2014, 2015, 2016 is like wars of magnitude more than you would have in ‘99, 2000.

Ameen: Absolutely. And I think that's a really good point you're bringing up there because I think it ties into this, the whole question, right?

Like, how do I, everyone wants to use machine learning now. Because the admin of the processing power made artificial neural networks and everything explode. And then people were like, we need machine learning scientists. And the machine learning scientists emerged and they were like, Hey, what about gradient boosting?

And there was all this data available and then we were like we wanna use the school neural network thing. So they build a neural net and they're like, okay, how do we deploy this where millions of customers on a website can actually hit an endpoint, which is powered by net? And the ML people were like I don't know your engineer.

And the engineer was like, I don't know what to do with this random python pickle. And that's that combination of all those things happening together, I think has made people realize that sometimes it's elegance to the simpler techniques. Like something as simple as a logistic regression can give you the answer rather than running into the woodworks of deep learning immediately.

As soon as, oh, we gotta use the latest and greatest neural network when a simple linear technique is sitting there, easily deployable with the computing power to just go wild over all the data you can possibly throw at it.

Bobby: Yeah it is like the confluence of a lot of these things, like you're saying, that kinda makes it come together.

And I think, companies like Nvidia are pushing GPUs hard for this reason, apparently based on the last quarter. Not hard enough, but but definitely that stuff is coming together. Looking into your crystal ball. And here we are towards the end of the year. And thinking about the trends ahead in 2019, what do you see as starting to gain some steam?

That was not really a thing on the radar, a year or two ago, but we’re starting to get into the machine learning thing.

Ameen: So I think, and we can come at this from two angles, right? One is the research, which is the cutting edge of machine learning. Where we've got, like you said, you've got companies like Nvidia, EMD like pushing the development of different chips and GPUs and you've got like GPUs becoming available as cloud computing resources in like various, like Amazon web services, you've got GPUs becoming available, more ubiquitous and cheaper to access. So I think we're gonna see an aggressive improvement in the use of deep learning on more like day-to-day problems, right? As people start using them more often. And it's easier to take out of box techniques and just run them right.

But I think more importantly, we are gonna see a kind of explosion in the area of auto ML, right? Everyone's talking about it now, where you've got a data set and you've got this library of different machine learning techniques, and one should you use for this data set to optimize for a certain why in the past when you needed like these unicorn machine learning scientists to be able to pull that off.

We've now got online cloud techniques that go in, not only find the right algorithm, but also hyper parameter optimization. It does all that stuff. And literally gives you an api, right? So it's like we're going to start unlocking problems that we didn't think were available to us for solving, because that barrier of requiring a mathematical savant or a statistical genius to be able to do your machine learning is gonna become an old one.

The question become problem formulation, like, how do we use machine learning to solve? Like we don't even know what we can use machine learning to solve. And we've seen this in the past with and I forget who said this, but like things like the internet, the television, airplanes, these are all things that people are like, oh, it's cool, but I don't know.

And I think we've got so much further to go with machine learning and I think the advent of this kind of. Automated techniques, the processing power of the chip, getting so much more ubiquitous that people are gonna start realizing that just the out of box techniques are ready for consumption, for solutions, for problems you don't even know you could solve in like smaller businesses.

While at the same times the larger research-oriented companies are gonna be driving forward on like the creation of art using AI and self-driving cars and flying taxis and this and that. So you're gonna have this development happening on both prongs right at time. And I'm excited to see that.

I'm excited to see the lower level one, where the simpler algorithms become more ubiquitous and day to day problems with more efficiency. It brings to mind—some mind to people said Excel was gonna kill financial analysts, right? Instead, what it did was it, you now have a million financial analysts that can all use Excel.

So I think that is what's going to happen with AI and ML as various auto ML tools kinda start kicking off and making machine learning more ubiquitous in the hands of people. And I think that, again, Excel quote, that Excel and financial analyst thing was another thing Pedro told me.

Bobby: I'll put it in the show notes and people can check that out. So your feeling is that we are still in early innings for what's possible with machine learning?

Ameen: Oh, I don't think we've even scratched the surface. I think we've found a shiny new toy and we are enjoying ourselves right now and we've got some big players that I think are starting to realize the critical value of ai.

But I think in the day to day, I think we're gonna get to a point where, you know, when you start a startup, what do you first get? You get a UX person, you get an engineer, and you start doing your thing, right? I think that AI and ML piece is gonna become one of those first pieces of the puzzle.

Like you don't need to wait before you can afford to get the ML scientist, or before you can collect data long enough, right? There's gonna be. I think open source databases, automated ML techniques, it's gonna become the cornerstone of problem formulation. And I don't think we've even started. And that being said, I think genetic algorithms as well have got, I think that's gonna be one of the next large family of algorithms that start getting utilized very aggressively for some very interesting problems.

And I'm very excited to see that happen as well. For sure.

Bobby: Yeah, if it uses the same principle of evolution that has a lot of legs, so exactly right.

Ameen: We seem to be doing alright, right?

Bobby: Yeah. We're ok. We're here.

No, this has been fantastic. So I'm just wondering, so when when your friends and family found out that you were going to be working at Zappos, did you become everyone's best buddy?

Ameen: Oh man. So at Zappos we can do 20% friends and family discounts, so I discovered friends I didn't know I had.

Bobby: Yeah like your friend Bobby, right?

Ameen: Absolutely! No, Zappos is it's a phenomenal place to work. Jokes side, actually. Absolutely love the people here. The culture is fantastic. Couldn't ask for a better team and a better group of people to explore ML and AI with. It's an absolutely fantastic group of people. We have here a bit of a shameless plug here.

Zappos is actually starting to venture out into. Sharing its expertise, I think it's expertise at Aappos.com that you can reach out to and be like, Hey, what are you guys doing? What are you up to? And we can chat with you there.

Bobby: That's fantastic. So we're gonna, we're gonna test your memory banks for trivia, but when you started at Zappos, do you remember the first thing you bought from Zappos?

Ameen: Oh man. So I, when I was an intern at Zappos was when I was first exposed to my employee discount. And I think the first thing I bought was a pair of Clarks. Chukka boots, if I recall. There you go. Directly. But I have, I, yeah, I could be right.

Bobby: I'm pretty sure that there, there's a, you could just go into your orders and find that out.

Ameen: Yeah, I can confirm, but I was being honest.

Bobby: That's terrific. This has been fantastic. I feel like this is the first of like a hundred episodes. I could do just on a lot of these topics, but I'm really respectful of your time. I know you gotta go and keep Zappos moving along. So I just wanted to thank you so much for taking the time and I really appreciate it.

Ameen: Absolutely. Thank you for having me. Happy to do this again. It was really enjoyable. I think it's a great discussion and I think we're both excited about the same thing, seeing where AI and ML’s gonna go in the future.So I'm sure we'll chat again as time goes on and we see it evolve.

Amim discusses his start in bioinformatics and how that provided a roadmap that took him to Zappos. It's yet another glowing example of how unpredictable and interesting one's career journey can be. Stick around until the end and find out the first thing amine ever bought on Zappos. This is the very first time we are doing the Loca podcast remotely.

Usually I have the good pleasure of being face-to-face with my guests, but in this particular instance, we decided to try something different, push the itches of technology and do this remotely. So while I'm in my office here in San Francisco, A mean is over at Zappos, I believe in Nevada. Is that right?

Ameen: Absolutely. Las [00:01:00] Vegas. All

Bobby: right. Las Vegas, Nevada. So without further ado, why don't I let my guests

Ameen: fully introduce. Hi, my name's Amin Uni. I currently lead the data Science and Data Science platform teams at Zappos. First of all, thanks Bobby for having me. This is very exciting. Enjoyed a few of the other episodes and I'm excited to be here and talk through data science.

A little bit about myself before we jump into the nitty of AI and machine learning. I started off doing my undergrad in just regular computer science, which is where. I was first exposed at the University of West Georgia. It's a tiny university in Carrollton, Georgia. It's got a fantastic CS program. And what was really exciting about it is the chair of the department at the time had this vague, like this interest in robotics and genetic algorithms, which surprisingly the genetic algorithms bit came back much later in the career, in my career, and I'm sure we'll get to that.

But I did a few independent studies with him where we built fun robotic [00:02:00] projects and. That's where my first brush with machine learning and AI happened as an undergrad. Over there we published some papers on using robots as tools to teach search algorithms and teaching robots how to walk like human beings using the Darwin anthropomorphic robot that they had back.

I don't know whether that's still at play in the robotic community, but. That's when I got first exposed to int intelligence systems and it was very exciting to me. But at the same time, both my parents are doctors, so there was still this little piece in of me that was like, I too will be a doctor, not a computer scientist.

That's what led me to get into this hybrid field in grad school with At Emory where I was studying biomedical informatics, it was pretty much a kind of detailed dive into using ML and good data practices to access data from various modalities like flow cytometry data, imaging data from MRIs, genomic data [00:03:00] pathology reports, which is just natural language from physicians.

It's a hard charge of every different type of data you can get access to. And then you gotta like figure out how to use it all together. So we did a lot of cool projects there. I was pursuing my PhD at the time. It was pretty intense. It was a good solid three years and it laid down the foundations for most of my academic ML background.

And Dr. Ashish Sharma, there was a fantastic guide to me and we worked together on a variety of different projects, which we can jump into a little more detail as we continue chatting, but it laid down the framework for being able to use different ML techniques for different types of data, and understanding how much diversity there was in machine learning as a field independent of the domain you're working in.

And I think that's what kind of sparked my interest too. Go out and do an internship at Zappos and get a feel for how could I use these same techniques we using here in a completely different domain. That's what got me super [00:04:00] excited. I saw an opportunity to get into this e-commerce space and see how we can use the same techniques that I've been practicing over the years in robotics and healthcare and see if I can port very similar methodologies to e-commerce and leverage ML cross domain without having to change the techniques.

And that's what led me to Bumblebee Analytics, which was a small startup in Bombay, in India, where we dealt predominantly with, e-comm clients trying to build big data infrastructures and things like that. And soon after that I came back to Zappos to start their digital experience data science team.

Happy to report, but since then, it's now one of the largest tech teams here. It's doing some pretty cool work. We've got some very exciting projects currently powering the customer experience using AI and ML on the website, but it's been quite the ride so far. And I guess that's the introduction. That's me in a nutshell.

And that's the. That's the various areas of ML and domains in which [00:05:00] I've slit it about in this field. Yeah,

Bobby: su super interesting and near and dear to my heart cause I'm spending more and more of my time and thus Locus time working on more projects in the life sciences space is something that's near and dear to my heart, in fact.

You should check out the previous episode of my podcast or more on that, which is pretty interesting. So what I was gonna ask was, so when you went to Emory, you were initially thinking about pursuing a PhD

Ameen: then, is that right? Yes. I was proud PhD dropout Over here. Right here.

Bobby: Yeah. No. Yeah.

And you're in fantastic company. It seems to be that's a requirement for successful people is having wherewithal to know when direction. You were three years into it. Yeah. So what made you decide to try a different direction at that point and not

Ameen: finish the PhD? So I think for me it was the base of it.

Something about the biomedical informatics. And this is not anything against research and academia, especially in the medical field. Cause I think it's [00:06:00] critical to be able to use things like ML and AI to drive healthcare forward. And I think it's going to be that differentiating factor that. Analogous, takes us to the moon and healthcare effectively.

But for me in particular, I liked the idea of seeing what I was doing immediately impact the domain that I was working in. And I think e-commerce, and that's very selfish of me honestly, but in e-commerce it gave me the ability to develop an algorithm and immediately see its impact on a customer base, like immediately affect the people.

I was trying to affect using machine learning and mentally, I think the shift from healthcare to shoes was a little, it was interesting for me to make that transition from healthcare to e-commerce in general. But I think it was the pace of it academia, especially in the medical field, requires discipline in terms of being patient, f d a approvals.

Clinical trials, long drawn out processes to see these things actually do what you intended [00:07:00] them to do. While in e-commerce I've had the opportunity to fail fast and, fail hard, but also have these successes immediately. Give me the gratification I was looking for while practicing machine learning.

That makes

Bobby: a lot of sense. Clearly the whole notion of failing and failing fast and healthcare has very different repercussions. But if you're an action oriented person that wants to constantly see improvement in results and get feedback so you get better, those feedback cycles are much more tenuous and longer.

I can definitely little.

Taking what you learned in this bioinformatics space and using it in different domains let's stay in the bioinformatics for a second and just for folks that dunno, maybe just take a couple of examples of the more common types of machine learning algorithms that are used, today in the bioinformatics life sciences space.

Maybe just taking a few examples that, that people could

Ameen: rock around. This takes me back to my first intro to biomedical informatics class that I took back in the day where the professor [00:08:00] that ended up becoming my advisor, Dr. Sharma, he gave me this example, which really captured, was very captivating to me in terms of what biomedical informatics was.

And he was giving me one of the first examples of biomedical informatics. And I forget the exact name of the town and hopefully I'm not getting the disease wrong, but there was a town in, I think England, where there was a cholera outbreak and nobody could figure out what was causing it. And what was amazing is that there was, I forget whether it was a physician or not, but the point was that somebody took a map of all the wells in the town and overlaid it on top of a map of where the chole outbreak was happening.

And was just like, this was an example of medical informatics being used for the first time to identify that these were contaminated wells in the city that were causing the disease. That example is what kind of laid the groundwork for what my research and stuff in the biomedical informatics space got informed by was [00:09:00] there are all these different modalities of data that are important.

By themselves, very important by themselves, but an absolute goldmine when you're able to look at them together and being able to view this cohort discovery as we called it, like being able to just query different modalities of data together, gave you a view into problems that you didn't know could be solved because you thought you didn't have the information.

I think that was fantastic for me in terms of machine learning algorithms in particular. In the life sciences space. It's so diverse, right? Just step stepping away from work that I was doing in particular, named entity recognition is a field of national language processing that uses an assortment of different techniques like conditional random fields or even LS dms and recurrent neural nets.

Now, to do. Named disease entity recognition in like pathology documents. And in addition to that, you can like completely switch gears and go to like image segmentation algorithms for [00:10:00] identification of cancerous nuclei and whole slide images. And you can go from there and switch gears again and actually look at using much more traditional like classification techniques like random forests.

To be able to make predictions on simple clinical data that you have available to you in just electronic medical records. So what was cool about it though is that almost every single one of those techniques works when you're trying to sell shoes as well. And that's what's really interesting to me about machine learning is how domain expertise really informs the algorithm you choose and how you use it.

But at the same time, you can take what you understood about one domain. And leverage it in another very, got some cool examples for that we can jump into? Yeah, why don't we, why don't we go ahead and do that? There's this cool analysis type called survival analysis. That they used to, I think it was called the West Nile Disease.

I think that's what the name of the disease is. But effectively what the West Nile virus and what they did [00:11:00] there was they used this survival analysis to show that people who were infected with it had a lower, a significantly lower chance of survival. Effectively showing. And the reason survival analysis is different from regular, just comparison of two cohorts is that you actually have, let's say during a clinical trial, people drop outta the clinical trial or they don't.

Do the things you expect them to do in the clinical trial. It helps account for those kind of anomalies where let's say if somebody disappears from a clinical trial, it doesn't necessarily mean they passed away. They could have just chosen to stop coming to the clinical trial. So how do you account for that?

So we were actually using machine learning to be able to help predict a customer's shoe size on the website. And that's a feature that's live on the website right now. Where we can guess your size accurately in whatever shoe you're gonna buy. If you've been a customer with us for a while, and the, what we used to evaluate if that was working or not was treating the return of a shoe as the analogy to like the shoe dying.

But someone [00:12:00] keeping the shoe did not imply that it fit them. It just imply that it was no longer part of the trial. So it could be that it fit or it could be that they left the trial and we actually used survival analysis, which is. A favorite of the CDC C for well survival analysis to figure out the survival rate of shoes to evaluate return rates.

So that was an example of how, and that was it. It gave us insight that we originally didn't have because we weren't just comparing two groups of people return versus not returned. It was like accounting for that. Maybe they just kept it and it didn't fit them. Fascinating.

Bobby: I never in my head would've been able to stitch that together in treating returns in, in, in the way that you did there.

That's so fascinating. And I'm guessing there are a lot of examples that are like that where you're thinking about healthcare and treatment and patient outcomes, and you can transform that into e-commerce metaphors.

Ameen: Yeah there's various ways to do that. What I think I found very helpful for me is and this is not maybe like a healthcare technique in [00:13:00] particular, but just genetic algorithms are something that we use very heavily.

Bobby: Actually, maybe just for the folks in the audience that don't know, maybe you could just take a step back and explain what genetic algorithms

Ameen: are. Oh, absolutely. Yeah. So a genetic algorithm effectively is a. It's a suite of, it's a, so they're known as evolutionary algorithms. They've got a flavor of randomness to them, and they evolve and adapt to the ecosystem that they're dealing with.

A genetic algorithm in particular is heavily influenced by the basic philosophies of evolutionary biology, which is that there's two principles that are used in the algorithm development. One is that the more fit an animal is, the more likely they are to reproduce. It's kinda like survival of the fittest, which.

Interesting point. Darvin did not say that. I think it was an economist, but besides the point. But yeah, the more fit a parent is, the more likely they're to reproduce. And the second principle is that traits of the parent carry forward into their children. So very simply put, those are the two core [00:14:00] principles that drive genetic algorithms and the fitness or how fit apparent is in the evolutionary senses, merely survival.

But in terms of an algorithm can be modified to be any kind of objective function. May it be increasing sales, may it be increasing conversion, may it be increasing search page clickthroughs. And you can use that fitness function to search a seemingly impossible to search large dimensional space to identify an optimal solution or a candidate solution that never occurred to you.

So the algorithm evolves adapting to the ecosystem and finds an optimal solution. That may not have occurred to you. You're not saying, oh, here's six options. Let me figure out which one's the best. All you are saying is, here's the problem. Here's three guesses. Find me a thousand solutions and figure out which one's the best.

And it like that evolutionary technique. So it's quite phenomenal to watch evolutionary biology at play. At solving a search algorithm.

Bobby: Okay. So would the, if you were trying to take that a step further and give an [00:15:00] example of using this genetic algorithm to solve a real world problem, would a search algorithm be where you would start or is there like a, another

Ameen: one that's, there's a much simpler example.

I think the search algorithm is something that we've actually been working on at Pure, and it's been, it's probably one of the hardest things we've tackled and it's been happy to say it worked, but I think a simpler example may be better for someone who is being the genetic algorithm for the first time.

So I think a lot of the listeners who are interested in AI ML are probably familiar and like just general optimization, are familiar with what's known as the traveling salesman or even the nap sat problem, right? Where you've got a bag that can hold 20 pounds and you've got 50 books, and the goal is to get the maximum number of books in this bag.

That, and the books have varying weights. So let's imagine a simple list of zeros and ones of 50 of them. One for each book, right? Every time. One of them's one instead of zero. It means the books in the bag. Now, if you try and think of every possible candidate [00:16:00] solution, it's a ginormous number of solutions as to all the combination of books together that are under 20 pounds that can fit in this bag, right?

And you can't possibly go through every solution. But a genetic algorithm will allow you to start with a random combination of books that are in the bag, maybe two or three candidate solutions, and then you start the evolutionary process and it figures out which book being in the bag is most useful for being able to fit more books in the bag.

So let's say there's a 15 pound book, right? Every candidate solution with that book in the bag is probably gonna fail because you can only fit books under five pounds in the bag. Now you've already taken up. 15 pounds with one of the books. So that book as a genetic trait is slowly going to die off right in the evolution of this solution.

All I can find that, oh, I've got 40 of these books at all. One pound each. Those are gonna be really strong genetic traits because you already got 40 books in that bag and you've still got 10 pounds to [00:17:00] play with, right? So those genetic traits are gonna keep propagating downward from population to population and probably end up in the fittest surviving solution.

Oh, interesting. And that's how, it's cool to think of it that way. Thinking of each book being in the bag as a genetic trait, like black hair or sure. Baldness and stuff like that. And these traits either carry forward or don't. And you can use this concept to optimize for any fitness function.

Bobby: So the NASDAQ problem is one that is easy to kinda understand cause it's famous. But in terms of and this can be outside of Zappos, but just in e-commerce in general, let's say, are there, more famous examples of how companies have utilized genetic algorithms, have you know, better outcomes for their business?

Ameen: I actually think that. We are one of the few people using it in an at scale solution. I've given a few talks about it and actually done another podcast focused purely in genetic algorithms. And I can tell you why I think they're useful in a situation like e-commerce, right? Any machine learning algorithm that you [00:18:00] build in e-commerce has to account for a changing customer base because you've got a mix of loyal customers, new customers, first time purchasers coming into your website.

Then you've got seasonality. It's cold. People want boots. It's summer. People don't want jackets. It's people want flip flops. And then you've got a constantly changing inventory, like what do you have in stock? What do you not have in stock? So a machine learning algorithm that you train at one point of time works for that state of your ecosystem, of your customer base, your seasonality for the domain you're in, the inventory position that you currently hold, right?

Now summer becomes winter. Your inventory changes. The customers you have either moved on or you've acquired new customers. And the general seasonality for whatever it is you're selling has changed. What do you need to do here? You need to build a brand new machine learning algorithm. You need to retrain it to account for that seasonality.

While if you've got a genetic algorithm that's constantly running and constantly evolving, it's tremendously valuable cuz it [00:19:00] adapts to the ecosystem. As evolution allows you to so it adapts to the ecosystem. And a great example there is a search algorithm like being able to surface the right brands and the right products to the right customer.

Given the current seasonality, given the current time of year, given the customer's current preferences and their most recent behavior is a valuable insight that a static machine learning algorithm cannot provide you.

Bobby: Yeah, so it can be, it's much more adaptive and therefore, the value of the query that it's hitting the mark of what the person, what the intent was, is much, much higher.

Cause you have a lot of moving variables. Yeah, that's pretty fascinating. So I wonder, I'm sure, I'm guessing Google probably uses a very end of this for their own queries. And so you could be the same guy and type in the same query, and if you did that 365 days a year, even from the same IP address, it may be a different result based on these other

Ameen: environmental factors.

Exactly. And I can obviously speak to what Sure yeah. I would imagine that I think that it definitely would be a candidate solution [00:20:00] for that kind of problem.

Bobby: Fascinating. Okay, so let's zoom out again and use kinda your fantastic experience on this to do another primer, which is, imagine being executive coming into this world and hearing and seeing headlines about artificial intelligence, machine learning, and deep learning.

Just those three terms. Okay. Ai, ml, and deep learning. How would you just describe at a high level how those things are connected and what the diff how they're related and so

Ameen: forth. Oh man, that is one of my favorite questions because it's asked so often, and I'm sure you've asked this question to other people as a, and my guess is you've always gotten different answers.

At least that's what, when I ask people that question, I always get different answers, right? So I'll give you my personal opinion here in terms of what the core differences are. So I think between ML and deep learning it's very simple, right? Deep learning is like a subset of ml. Your logistic regressions, your XG boost, your random forest.

Your clustering is all machine learning, right? So machine learning is broken up in a supervised [00:21:00] and unsupervised learning. And now you've got reinforcement learning as well. Supervised learning is effectively when you've got historical data that, and you have tagged historical data. So let's say I'm trying to use all of someone's click-through behavior to predict if they're gonna convert on the website or not.

I've got historical indicators of what people have done on the website and whether they convert it or not. So I can train a model on that and then predict on unseen data points when someone's live on the website. And see if they're gonna convert or not. So that's supervised learning. Our supervised learning is when you're like doing things like clustering.

So algorithms that fall into this family is ca means clustering, hierarchical clustering. These are techniques that basically allow you to create groups within your data points without actually having a label associated with it. So an example there is, if I told you, oh, I have something that's, Yellow and long, it's probably a banana.

But if you didn't have the tags and you just had [00:22:00] a bunch of fruits and you try to cluster them, you'd probably cluster based off of color, taste, shape, and create like subgroups and you'd be able to like assign hopefully names to them. Or if you're using cars, you could use different principles and properties of these cars and get oh, these are oh vehicles you could get, these are airplanes, these are boats.

These are cars. These are trucks, these are sedans. These are hatchbacks, and this is a fighter jet, a commercial airline, a speedboat, streamliner. And these kind of become like this is known as hierarchical clustering. You keep breaking it down, right? This is when you don't have a label. And reinforcement learning is a little less known of than your classical supervisor and unsupervised learning.

But it's effectively when you've got like semi-supervised data, right? You've got some label data, you've got a human actor informing the algorithm as to whether it's right or wrong, and it just keeps learning from that. So this is the family of machine learning, right? Deep learning or artificial neural networks and the variety of artificial neural networks that are [00:23:00] out there.

Generative adversarial networks, recurrent neural networks, long short term memory, ls, dms, et cetera, et cetera. Are all the hotness in machine learning right now? Those are examples of supervised machine learning techniques, right? They have historical data, things that they've, and they have a label, and you're able to learn from that historical data to be able to make a prediction on something you've not seen before.

New data points. These deep learning techniques are, they're more black box. They've got multiple layers in the neural network layered together. It's not as, it may be not as interpretable because you don't know what's going on inside the various layers of the neural network, but it's basically like using all the various combinations of your input data to make a black box kind of prediction.

It's able to learn patterns in relationships that you wouldn't normally be able to do in a kind of a non-linear technique. Now AI is where it gets a little difficult to come up with a definition. I've heard a variety of definitions. One being where AI is any kind of [00:24:00] artificial intelligence, right?

There's this fantastic data scientist, his name is Pedro Alvez. I worked with him a lot in the past, and he gives an example of ai, which is a paper towel dispenser, right? It's just mimicking intelligence, right? Like when, if I asked you for a paper towel right now with my hand, you'd give one right? Go to a paper towel dispenser in a bathroom and you stick your hand underneath it and it like spits out a paper towel that shows some semblance of intelligence, right?

Technically that's artificial intelligence right there. Or you can see AI as like the general AI where everyone's oh, the singularity, and we are gonna get machines that can act like human beings. So literally completely two opposite ends of the spectrum can be used as a definition of ai. I personally believe that.

The theory behind machine learning and AI came long before the hardware that allowed us to do AI and machine learning. Back when that had happened, they had called it AI and then it fizzled out because it got super [00:25:00] exciting and then no one could do anything with it cuz the chips didn't exist to do it.

So what happens when you've got technology that you can't use? You get sci-fi movies and sci-fi movies aren't like, oh, look at this AI in this beautiful world and just help. It's the Terminator, right? So you've gotta buy, you got a lot of bad press associated with ai, and then suddenly we got the chips to be able to actually do the neural networks and stuff.

So we just rebranded it as machine learning and did it anywhere.

Bobby: And it became, yeah, and it, and the rest is history as they say. Per year prediction, I think I get asked it and I ask this question quite a bit more and more these days, and I definitely do get different answers, but that's definitely one of the better ones I've heard that resonate with a lot of my own feelings, so That's great.

I appreciate you taking the time to do that. So if we then dive a level deeper and start talking about, just basic algorithms and you. I'm just curious on your thoughts. If you take something like gradient boosting, that's something that's I think has been around since the late nineties, but has been picking [00:26:00] up steam much more recently.

Like to what do you attribute that, decades gap in between, discovery and mainstream

Ameen: use? So I think this is, again, answer that I think is a personal opinion and we do a lot of gradient boosting as well. It's basically optimizing a cost, right? Gradient boosting is effectively optimizing some cost and it's like trying to like maximize something, right?

So very simply put like a good example of gradient boosting is like tree boosting, right? You've got a lot of different gradient tree boosting. It's used for like classification, regression, et cetera. The advantage that I've found there is twofold, right? One, it's interpretable. I think as ML has picked up, there's a lot of companies that do a lot of research in AI and ml.

But you'll find that a lot of companies have data and they don't really know what to do with it, right? Like they have a lot of data, but they don't have the requisite talent. And even if they did get the requisite talent, it's very difficult to deploy neural networks and these deep learning techniques in [00:27:00] production and environments, right?

So if you go into, you look at Zappos as an example, you look at any other retailer as an example, or you just go and look at the field as a whole. And you or you run a Google search for, how do I deploy a long short-term memory network at scale in a Java api? It's really hard to do that. There isn't like stock packages out there, and I give a talk on this on machine learning in production, which talks about a few ways to navigate around this problem, but there's not a lot of well defined methodologies out there, but gradient boosting or the gradient tree boosting provides you with solutions that can be represented in.

Standardized formats like predictive model markup language pmml, which allows you to use J P M ML in a standard spring boot application where API developers are familiar with what they're doing. You don't need machine learning engineers writing your production API code. They're able to take the knowledge of your machine learning engineers and put it in a production setting.

You're no longer looking for those unicorns who are like at [00:28:00] scale software engineers and deep learning specialists at the same time, and more importantly, you understand the why. In a gradient boosted algorithm, you're able to actually know what variables are influencing your output and how, and it's not just a black box technique that you need to rely on where there's a lot of research being done in increasing the interpretability of neural networks right now, but it's not quite as simple as, here's a tree if it's greater than five for this variable, go this way, otherwise go that way.

The and ease of deployment I think are. Critical in making them so popular now that people are starting to ask the question, how do I use machine learning in production? So

Bobby: those the, the accessibility to it was something that was pretty true in the late nineties. But I think the difference, building on what you're saying, it, it seems to me that I think there were these two big.

Step function changes in the power of computing that then made them become popular in the last couple years. Something like grad boosting, like why did gradient boosting not become like [00:29:00] a known term in the late nineties versus like now? And I think it's one, the processing power of chips, which you alluded to earlier, was not there.

And the second thing I think is just the explosion of data itself, right? The amount of data that is out there in 20 14, 20 15, 20 16 is like wars of magnitude more than you would have in 99,

Ameen: 2000. Absolutely. And I think that's a really good point you're bringing up there because I think it ties into this, the whole question, right?

Like, how do I, everyone wants to use machine learning now. Because the admin of the processing power made artificial neural networks and everything explode. And then people were like, we need machine learning scientists. And the machine learning scientists emerged and they were like, Hey, what about gradient boosting?

And there was all this data available and then we were like we wanna use the school neural network thing. So they build a neural net and they're like, okay, how do we deploy this where millions of customers on a website can actually hit an endpoint, which is powered by net? And the ML people were like I don't know your engineer.

And the engineer was like, I don't know what to do with this random python pickle. And that's that combination of all [00:30:00] those things happening together, I think has made people realize that sometimes it's elegance to the simpler techniques. Like something as simple as a logistic regression can give you the answer rather than running into the woodworks of deep learning immediately.

As soon as, oh, we gotta use the latest and greatest neural network when a simple linear technique is sitting there, easily deployable with the computing power to just go wild over all the data you can possibly throw at it. Yeah it is like

Bobby: the confluence of a lot of these things, like you're saying, that kinda makes it come together.

And I think, companies like Nvidia are, pushing GPUs hard for this reason, apparently based on the last quarter. Not hard enough, but but definitely that stuff is coming together. Looking into your crystal ball. And here we are towards the end of the year. And thinking about the trends ahead in 2019, what do you see as starting to gain some steam?

That was not really a thing on the radar, a year or two ago, but starting to get a. In the

Ameen: machine learning. So I think, and we [00:31:00] can come at this from two angles, right? One is the research, which is the cutting edge of machine learning. Where we've got, like you said, you've got companies like Nvidia, emd like pushing the development of different chips and GPUs and like you, and you've got like GPUs becoming available as cloud computing resources.

In like various, like Amazon web services, you've got GPUs becoming available, more ubiquitous and cheaper to access. So I think we're gonna see an aggressive improvement in the use of deep learning on more like day-to-day problems, right? As people start using them more often. And it's easier to take out of box techniques and just run them right.

But I think more importantly, we are gonna see a. Kind of explosion in the area of auto ml, right? Everyone's talking about it now, where you've got a data set and you've got this library of different machine learning techniques, and one should you use for this data set to optimize for a certain why in the past when you needed [00:32:00] like these unicorn machine learning scientists to be able to pull that off.

We've now got online cloud techniques that go in, not only find the right algorithm, but also hyper parameter optimization. It does all that stuff. And literally gives you an api, right? So it's like we're going to start unlocking problems that we didn't think were available to us for solving, because that barrier of requiring a mathematical savant or a statistical genius to be able to do your machine learning is gonna become an old one.

The question become problem formulation, like, how do we use machine learning? To solve. Like we don't even know what we can use machine learning to solve. And we've seen this in the past with and I forget who said this, but like things like the internet, the television, airplanes, these are all things that people are like, oh, it's cool, but I don't know.

And I think we've got so much further to go with machine learning and I think the advent of this kind of. Automated techniques, the processing power of the chip, getting [00:33:00] so much more ubiquitous that people are gonna start realizing that just the out of box techniques are ready for consumption, for solutions, for problems you don't even know you could solve in like smaller businesses.

While at the same times the larger research-oriented companies are gonna be driving forward on like the creation of art using AI and self-driving cars and flying taxis and this and that. So you're gonna have this development happening on both prongs right at time. And I think it's, I'm excited to see that.

I'm excited to see the lower level one, where the simpler algorithms become more ubiquitous and day to day problems with more efficiency. PE it brings some mind to people said Excel was gonna kill financial analysts, right? Instead, what it did was it, you now have a million financial analysts that can all use Excel.

So I think that is what's going to happen with AI and ML as various auto ML tools kinda start kicking off and making machine learning more ubiquitous in the hands of people. And I think that, again, Excel quote, that [00:34:00] Excel and financial analyst thing was another thing Pedro told me.

So Pedro Al, you guys should check. Yeah,

Bobby: I'll put it in the show notes and people can check that out. So your feeling is that we are still in early innings for what's possible with machine learning?

Ameen: Oh, I don't think we've even scratched the surface. I think we've found a shiny new toy and we are enjoying ourselves right now and we've got some big players that I think are starting to realize the critical value of ai.

But I think in the day to day, I think we're gonna get to a point where, you know, when you start a startup, what do you first get? You get a UX person, you get an engineer, and you start doing your thing, right? I think that AI and ML piece is gonna become one of those first pieces of the puzzle.

Like you don't need to wait before you can afford to get the ML scientist, or before you can collect data long enough, right? There's gonna be. I think open source databases, automated ML techniques, it's gonna become the cornerstone of problem [00:35:00] formulation. And I don't think we've even started. And that being said, I think genetic algorithms as well have got, I think that's gonna be one of the next large family of algorithms that start getting utilized very aggressively for some very interesting problems.

And I'm very excited to see that happen as well. For sure.

Bobby: Yeah, if it uses the same principle of evolution that has a lot of legs, so Exactly right.

Ameen: We seem to be doing alright, right? Yeah. We're ok. We're here. We more podcasts now,

Bobby: How did we get the technology to do this?

No, this has been fantastic. So I'm just wondering, so when when your friends and family found out that you were going to be working at Zappos, did you become everyone's best buddy?

Ameen: Oh man. So at Zappos we can do 20% friends and family discounts, so I discovered friends. I didn't know I had Yeah like your friend Bobby, right?

Like

Bobby: you

Ameen: Absolutely No, Zappos is it's a phenomenal place to work. Jokes side, actually. Absolutely love the people here. The culture is fantastic. Couldn't ask for a better team and a better group of people to explore [00:36:00] ML and AI with. It's an absolutely fantastic group of people. We have here a bit of a shameless plug here.

Zappos is actually starting to venture out into. Sharing its expertise, I think it's expertise@zappos.com that you can reach out to and be like, Hey, what are you guys doing? What are you up to? And we can chat with you there.

Bobby: That's fantastic. So we're gonna, we're gonna test your memory banks for trivia, but when you started at Zappos, do you remember the first thing you bought from

Ameen: Zappos?

Oh man. So I, when I was an intern at Zappos was when I was first exposed to my employee discount. And I think the first thing I bought was a pair of Clarks. Chaka boots, if I recall. There you go. Directly. But I have, I, yeah, I could be right.

Bobby: I'm pretty sure that there, there's a, you could just go into your orders and find

Ameen: that out.

Yeah, I can confirm, but I was being honest in our

Bobby: That's terrific. This has been fantastic. I feel like this is the first of like a hundred episodes. I could do just on a lot of these topics, but I'm really respectful of your time. I know you gotta go and keep Zappos [00:37:00] moving along.

Yeah. So I just wanted to thank you so much for taking the time and I really appreciate it. Thank you so

Ameen: much. Absolutely. Thank you for having me. Happy to do this again. It was really enjoyable. I think it's a great discussion and I think we're both excited about the same thing, seeing where AI n ml's gonna go in the future.

So I'm sure we'll chat again as time goes on and we see it evolve.