Getting Started with Deep Learning in the Enterprise

Dave Cole

Hello! Welcome to another episode of the Data Science Leaders podcast. I’m your host, Dave Cole, and today our guest is Eitan Anzenberg. He’s the Chief Data Scientist at Bill.com. He also has a PhD in physics from Boston University. And you majored in astrophysics at UC Santa Cruz in your undergrad. Quite the profile. Eitan, welcome to the Data Science Leaders podcast.

Eitan Anzenberg

Yeah, thank you, Dave. Thank you for having me. I appreciate it.

Dave Cole

Awesome. So today we’re going to be talking about deep learning, so getting started with deep learning. I know that’s something that you have brought to your current role at Bill.com, talking a little bit about that journey and helping others get started, and maybe talk at high level about types of use cases, where to start, what to expect, all that good stuff.

The next thing is, there are a lot of PhDs who are data science leaders who come on to the DSL podcast. But not all PhDs are in physics. So, I call that a somewhat unrelated field. We’ll talk a little bit about what’s translated, what hasn’t translated from your PhD and also, is there an untapped market in hiring physics PhDs and bringing them into the world of data science?

Eitan Anzenberg

It’s possible. I mean, they’re not cheap, so sure, it’s always possible. But yeah, my journey has been somewhat unique. I found from data scientists, machine learning researchers that I’ve spoken to, generally, I see a lot of unique journeys. There’s no one way to do things, which is really cool, actually, which attracts me a lot to this space, because we see varied thoughts and varied ways of tackling very challenging problems in industry. Quite different in terms of challenging problems in academia.

I came from the academic world, a lot of stuff has translated over. I would say pretty much any math statistics that we were dealing with, when I was taking data as a physicist we were dealing with small data sets, and when you have small data sets, statistics is super important. Now in machine learning we deal with huge data sets, the statistics are still actually incredibly important but in different ways in terms of bias, but also just thinking about problems holistically, thinking about how to tackle challenging problems.

In physics the joke is like every cow is a sphere. How do we break down to the simplest form of something and build a model — I mean, it’s by paper, by hand, but how do you build a model that tries to encompass what you’re seeing in the world? You’re trying to model the world, in this case, in deep learning…

But machine learning, in general, we’re modeling data. Nowadays because of compute power our data can be quite large. You have video, you have audio, pictures, texts, a whole corpus of Wikipedia or Yelp or whatever you’re dealing with. So we’re dealing with big, big amounts of data, and you’re trying to get understanding, trying to get even just some insight from that data. Deep learning in these kinds of domains is actually very powerful.

It’s a tool. I like to say it’s a tool that works really well on certain domains, and then in other domains we still use classical machine learning, like random forests, XGBoost, LightGBM, even AutoML these days. Maybe that’s one thing that’s changed quite a bit, is I’d be more inclined to use AutoML, which I don’t know if your audience will be familiar with. But it’s essentially, you have data, that’s always like the beginning, the first point, trust your data.

You want to model it, but you’re not going to do the hyper parameter searching, you’re not going to try to optimize the model. You have an algorithm that tries to optimize what the model should look like, so whether it’s an ensemble of 10 different models of all these different hyper parameters. And it’s doing that through statistics. It’s doing it also because we have compute power. So AutoML, just in my experience now more recently, is very powerful.

Whenever I think about maybe building a scikit model I also tend to try to do an AutoML model, which will have an ensemble of a bunch of different models. But deep learning, it’s beyond AutoML because of the compute. For me sometimes it takes a week to train one model, and that’s on one GPU, and the GPU is already fairly expensive compared to CPU. And so you’re not going to build a thousand of these and try to hyper parameter search. Also, the parameter space is way, way large. It’s crazy big compared to an XGBoost model. Where to begin, where to begin with deep learning?

Dave Cole

Hold on. Let me slow you down. There’s a lot of great information that you’re thrown at me here. So one of the things I heard you say, and I don’t want to move on from it until we unpack it a little bit, is that you’ve become more enamored with the power of AutoML, so the ability to basically throw data at an AutoML package and have it come up with the model and optimize for whatever dependent variable you’re trying to predict or what have you. Are you saying that you’re leveraging AutoML as a way of comparing it against your traditional scikit-learn model that you come up with on its own? It’s almost like a benchmarking exercise?

Eitan Anzenberg

Yeah, absolutely. Yeah.

Dave Cole

And then how much are you trusting an AutoML model? So are you trusting it no differently than you would, say, a model that your team or you built using scikit-learn, or no?

Eitan Anzenberg

Yeah, I would trust it as equal if I consider this, so you have the same training data and you have the same validation data, a holdout that you test at the very, very end, after you’ve done all your mixing and matching. Within your training data, you’re going to split it a million ways. You do your random split. And that’s how AutoML works, it’s tuning hyper parameters as a function of your test set performance. Your test set’s coming from the training data, but every new iteration is a new slice of the train data.

In terms of how I trust it, at the end of the day I trust it as well as I trust my data, what the data is. When I talk with my team or when I talk with people that are building the models I always start with, Do you trust the data? Is it the universe of what you’re trying to model? And I mean it in that scale.

I give talks as well about bias in machine learning or in AI. Bias is huge because if you have training data or data that’s a small subset of your universe and you build a model for it, your model is going to work okay in that local area of the universe, and when you try to expand it… I start to lose trust, I start to become a skeptic. So if I trust the data, of course, I’m going to trust the AutoML. It’s just an algorithm.

Even a random forest is an ensemble of decision trees. I mean, it’s not AutoML, but it is an algorithm that’s training a model. But that model is like 100 little models inside. You know what I mean?

So, ultimately, of course, I trust it, but the data is always important. I found successful data scientists and machine learning engineers or researchers, or whatever the names are of these teams, the ones that are successful are the ones that really try to understand the data. They try to dig in, understand what the use case is, how this is going to be used, what the data is that they’re trying to model. Is it holistic, is it the volume of the universe they’re interested in?

Dave Cole

Right. So let me stop you there. When you say, is the data that they’re working with the universe that they’re working with, are you basically saying, Hey, what we’re trying to predict, I want to make sure I have all of the possible features, all the possible data that is out there and that can be used at my disposal? If I’m missing something, it doesn’t matter how great the model is, it’s probably not going to perform as well if I’m missing some key features.

So, your role as a data science leader, how do you instill that mentality in your data scientists to make sure that they know all of the data that they need to solve a particular problem? Do you have a review stage before you dive into a project to make sure that they have all the data before they just go diving off and creating their model? What do you do to guard against that?

Eitan Anzenberg

Yeah, for critical stuff, like critical things in terms of bias, there’s protected groups, we’re in the US. In most of the world now there’s just protected groups. So if you’re modeling, as an example, on an age range that is, let’s say 18 to 34, I’m not sure I’m going to trust the model outside of that age range.

And so you can’t really apply that. It doesn’t have to be a machine learning model, it can be an algorithm, it can be a handwritten decision tree. But in terms of instilling, I tend to sound like a broken record around the office, I’m always like, “Do you trust the data?” That’s the number one thing. And there’s a lot packed into that statement.

Trusting the data means whatever you build the model comes from that data. You’re trying to represent the data. You’re not reproducing data, but you’re building a model on that data set. If you don’t trust the data or if it’s a biased sample, you wouldn’t feel comfortable applying it outside of scope.

The challenge is, the Holy Grail or when I found machine learning to be very powerful, is when it generalizes outside of your data set. So I’ve built models where, as an example, let’s say build a model on invoices in Great Britain, and then I’m not sure if it’s going to work for Singaporean documents that are also in English. But if it works, maybe the accuracy is not as high, but I feel like, Oh, that’s a really cool thing, because it does generalize. The accuracy is not that great, but ultimately…

So it’s like you want to be able to build models that generalize outside of scope. It’s almost like a Catch-22, you want your data to be a holistic view. So, ideally, you would train on data for your universe that you’re interested in. But, yeah, so it is a challenge of trying to balance those two areas. You want to be able to generalize, but you also want to have a model that is unbiased.

Dave Cole

Right. So, let me paraphrase a little bit here, so if you’re working with your team and they’re building a model, where the training data set is, a set of images of invoices based in the UK. You might ask them, Hey, I think we might be expanding beyond just the UK. The UK is just like our first project, and I do want you to make this model ideally a little bit more generic. We don’t know how the data set may be for other areas, but just think about it in a more generic way and try not to cut corners, and make it so specific to what you see in the UK.

There’s that balance there. You only get so general that you’re ready for documents and all sorts of different languages that are outside of English, and the scope creep can be quite crazy at that point. But, yeah, that balance is what I think you’re pointing to is one of your roles as a data science leader, is to help your data scientists think in those terms.

Eitan Anzenberg

Yeah. And in terms of committees there’s a book on extreme ownership that is around the office. I forget… like a couple Navy SEALs or something, but the point is that if you’re building it, you should own it. You can’t just depend on a committee to approve, even if it’s a sensitive thing. Hypothetically, let’s say you’re at a bank and you’re building models for credit lending or credit decisioning or loans, and you’re tasked to build a machine learning model for that, you need to essentially own that whole piece. So that if and when it gets deployed, that you trust the result, so you have to verify, you have to test it in all those extreme cases. Bill.com, we’re in FinTech, so anything finance related is heavily regulated. So there’s a lot of testing. We go through tons and tons of testing, but at the end of the day it falls down to the engineer, the researcher that built the model. If something doesn’t work, it’s on them to resolve it, to fix it, figure out why, and move on.

Dave Cole

Right. And when you say testing, are you talking about testing the accuracy of the model, like throwing different test datasets, or are you also talking about like, once the model is deployed, making sure it’s tested and it is operating within whatever SLAs you have and that kind of thing, or probably both?

Eitan Anzenberg

Yeah, both. Definitely both. And testing as well, if it’s related to bias in the marketplace… you see it in the news all the time. I mean, Apple’s, I think one of their credit cards as well, had different limits for the different spouses in the household, these types of things.

I hate to see it because I’m fairly optimistic. I feel like the people that released it had the best intentions in mind, maybe it just wasn’t tested enough. In FinTech, you have to cover your bases. Then when it’s in production, there’s a common phenomenon of just decay. So your model looks really good for days or maybe the first week, but a lot of models do decay over time because of the time dependence.

So you have to monitor. You have to monitor once it’s released. There’s a whole lifecycle of machine learning. It’s like once you release it, you’re not done yet, you still have to monitor, you have to make sure it still works. The SLA is incredibly important if it’s real time predictions. If you’re in the cloud, for example, there’s SLA, yes, so there’s a whole life cycle. It’s very interesting.

Dave Cole

I do want to get to deep learning, but I’m just fascinated. You’ve mentioned that you’ve spoken in the past about bias, and that bias, clearly, in FinTech is one of the big challenges. What is some advice that you would give other data science leaders who might be in a regulated industry and from a high level in terms of trying to limit, reduce, get rid of to what extent possible bias in our models?

Eitan Anzenberg

They probably already know this, the more interpretable the model, the better. And even going all the way to just not even doing machine learning, have an algorithm. Have an algorithm for your credit decisioning or for your fraud detection or whatever…

If your group is risk averse in this regard… like a lot of tech companies want to move fast and break things even in FinTech. They find that that’s their ultimate passion. That they just want to really move fast. But I would say that if you’re not comfortable with moving fast, start with an algorithm that you can see, that you know the limits of. So it’s a formula. Move from there to completely interpretable models, those are linear models or decision tree type models, so you can still build that off data, but you get an artifact that you can see. You can understand, Oh, this is the decision, this is why this applicant was accepted. You can see the logic that the tree went down.

Once you get into nonlinear models there… I know banks use things like explainability. I’ve actually given talks on explainability, so trying to understand nonlinear models and what they would call blackbox models. It’s just a nonlinear model. It has a curvature. So you can find a local area of that model that you can fully explain. It’s an approximation, but it becomes linear because you’ve zoomed into that local area. I mean, it depends how risk averse you are in your group and what the goals are. But I imagine a lot of people think that as well. If they’re in these heavily regulated industries, where they’re applying machine learning.

Dave Cole

Right. And that’s a good segue, I think, into getting started with deep learning. One of the challenges of deep learning is that black box moniker that it’s given. So segue into deep learning and talk a little bit more about getting started there, and what are some of the traps you have to be aware of, what are some ways in which to get started?

Eitan Anzenberg

The way that I got into deep learning is I was doing machine learning and I was reading, watching as much stuff, like Stanford has their master level classes on YouTube. But one thing I noticed coming from physics is how actually accessible deep learning feels. In physics, if you have a lab as an experimentalist the equipment costs are a million or above a million, if you’re doing big experiments it’s in the billions, obviously, but forget that.

But deep learning, as an example, I have a GPU machine, I threw Linux on it, and I’m able to train at night. In some ways, if you come from an academic science perspective, it feels very approachable in terms of the hardware. And then what you can do is I do a personal project, if it’s something you’re interested in, something you’re passionate about learning.

Ultimately, deep learning is a subset of machine learning. So I would start with machine. If you don’t know either, start with machine learning, you can train that on your laptop, probably even on your phone these days and your iPad if you have some data. Learn the statistics or the math that’s involved with some of the modeling, the testing.

Black box to me, I mean, random forest is a black box, XGBoost is a black box. A deep learning model is a bigger model. It’s sometimes a slower model, so it might take on my laptop a second to score, whereas a random forest might take like a tenth of a second. So, that makes it a little bit harder to work with, but you can apply similar techniques, so you can apply the same kind of explainability techniques to it.

And when you’re picking a personal project, I think we talked about some of the domains that deep learning has been very powerful in, so video, photos, or images, audio, text or language processing, there’s a lot of open data. So ImageNet for images, there’s like 100 million images. For text there’s Wikipedia I mean, the corpus, there’s, I think, even Yelp comments. There might be some other open datasets.

You’ll notice the size of the datasets, and it’s critical, because mentioning deep learning models, they’re big, and so to train them, you need a lot of data. You can’t train it from scratch on a small amount of data. But there’s techniques like transfer learning. So with transfer learning you take a pre train model and you fine tune it just a little bit on your maybe smaller data set. You can run it on GPU overnight and get some results the next morning. That’s stuff that I’m excited about doing.

But whatever project you might be passionate about or domain you might be passionate about, and if you are interested in learning more about deep learning or machine learning, just go. If you can only do it as a side project… At Bill.com, I don’t want to brag, but we’re very open to doing machine learning, but I know there are some companies out there that aren’t so open.

And if you’re starting out, if you’re starting in the field you may not have the opportunity to do it in your current position if you’re a software engineer or a data analyst, but don’t let that stop you. In some ways, if I could do it on a laptop, most people can do it.

Dave Cole

Yeah. My teenage son is trying to convince me that he wants to get a machine, like a PC type machine, and buy a GPU card because he wants to mine Bitcoin. So maybe I can convince him instead to use that GPU to build a deep learning model. I think that might be better for his long term career.

Eitan Anzenberg

Well, it depends on the coin. If it’s dogecoin, maybe. No, I agree. I mean, to be honest, when I was getting the machine it wasn’t just for deep learning. I like playing video games, and I’ve had Nvidia cards since I was like 14 or something. And this is way before any of this. But it’s just the passion. It’s cool to be able to train stuff overnight, spin it up, dig into CUDA, dig into TensorFlow.

When you’re in the cloud, it’s so nice. In the cloud everything is there. Everything’s pre-packaged. When you have your own Linux machine and you have to debug, Oh, why did my audio drivers cut out, why did my GPU… I can’t see it in the terminal. So there’s a sort of tinkering. It’s cool. I like that, digging into that level, because maybe you’ll miss some of that if you’re just always in the cloud and everything looks great.

Dave Cole

Right. Yeah, I know dealing with GPU drivers can be a real hassle. So, the advent of the cloud and having, whether it be Docker images or what have you, that are already ready to go just makes life so much simpler. That segues into a conversation about the cloud. So, what I’ve heard so far is if you want to get involved in deep learning, obviously, you should have some background in machine learning, and find a use case. I mean, was there a specific use case that you started with a hello world of deep learning? Was it some image classifier or something like that? Do you recall?

Eitan Anzenberg

I’m trying to remember. I think, actually, it was like a stock predictor, so like trying to use LSTMs. And I bought some stock data for the US markets. It was maybe in the tens of dollars. I think that was one of the first ones I attempted. I have played around with imaging, but I think actually within the roles that I’ve been in, and then, obviously, pre-train models are so powerful, because you can just then fine-tune and you actually don’t need that much data, but.

Dave Cole

Right. So I think what I’m also hearing is I need to come to you for stock advice as well is probably… I’m sure you’re at version 27 of that model predicting.

Eitan Anzenberg

This is not financial advice. I’m not running any models.

Dave Cole

That’s what this podcast is all about. But, yeah, you killed two birds with one stone on that one. So, at any rate, to start with something fairly simple, the advent of the cloud has been very helpful. Not everyone is going to buy a machine with a Nvidia GPU, and maybe you can help us too. Why are GPUs so helpful?Do you know why they’re so helpful when it comes to deep learning? Help demystify that for our audience here.

Eitan Anzenberg

Yeah. That’s a good question. I think that’s related to the parallelization of the… What are they called? There’s a core in a GPU processor. There’s a good distinction compared to… a GPU card to a CPU. There’s a notion of these core… I forget the name of the core, but it’s highly parallelizable. The way that it does math for constructing, a 3D scene is a lot like a corollary to the math involved with all the matrix math within doing your forward pass, backward pass to compute the gradient descent, make the updates, get your next mini batch, forward pass, backward pass. There’s a lot of parallelization within those steps. So I think that’s why.

And TPUs are the ones that Google are working on, I’m even less familiar with. There’s more specialized hardware. But I will say this for the cloud in general, this is a hobby. My desktop is a hobby. When you’re talking about building something that millions of people are going to use at scale at any time of the night, being able to spin up, spin down instances, monitor, logging, SLAs, and being able to revert things. That’s just the life cycle. That’s the part of the model where it’s running. And then there’s all the training.

So, when you have a team, we have a team that is using GPU resources, and you have a fixed number of GPUs, let’s say, in a closet somewhere, there’s going to be a queue of jobs to send at. Actually, many years ago this was how we were doing things. This was actually in an academic lab. And there’s someone that at the time they were maintaining a closet for the hardware for a bunch of scientists, and you would send jobs…

It’s a limited amount, obviously, to purchase more you have to wait, and it takes time. You have to plan for it. In the cloud, like if one day I need 100 GPUs for something, I can do that, or maybe I need one. Or if I’m running in production, yeah, like there’s big spikes all of a sudden, we don’t understand why, a lot of requests, they manage that, they manage that load balance. So it’s pretty incredible just how easy it is these days to do some of these.

Dave Cole

I mean, let’s talk a little bit about some of the, maybe the challenges that you faced, I mean, it’s not cheap, either.

Eitan Anzenberg

Yes, exactly. That’s the downside. The downside is it is more expensive. In some ways, maybe that’s a good thing. So if people are using it, if you see users enjoy it because either they’re getting results faster, more accurately, results that you couldn’t get with a classical ML model. So if they’re using it you have to purchase the hardware to represent that. But maybe that’s a good thing, because you have users that are enjoying the experience. That’s absolutely true, so in terms of the cost, but there are ways like… With AWS they have spot instances.

And then the other thing, and we go back to regulations, some banks — and they’re pretty open about it — they do everything on prem, they’re not comfortable yet. Some of the biggest banks in the world are still on prem. Eventually they’ll probably migrate to whichever cloud they feel comfortable with, but today they’re not. So it’s a journey for a lot of different established entities out there.

Dave Cole

Yeah. Are there any other pieces of advice or things to be aware of when starting your journey in deep learning? Obviously, start with learning, step two about deep learning, and then step two is really uncover a valid use case that makes sense. You don’t want to be throwing deep learning as a solution to every type of use case. I mean, you mentioned a few in terms of video, photo, text, classifiers, that kind of thing.

Did a member of your team spin up a GPU machine and use a deep learning type model for a use case that wasn’t remotely appropriate, but just for fun or something, racked up a huge bill? Anything to be aware of?

Eitan Anzenberg

It doesn’t happen too often, but I understand the sentiment. We have a really strong team, and they use the right tool for the job, and that’s really great, but I get the sentiment of wanting to try something new. By the way, you can train in neural nets in scikit. There are feed forward networks in scikit. Now, you won’t get the fancy convolution nets, and you can’t run on GPU, but it will train on your CPU.

But I will say that… all the domains I mentioned, I’m sure I missed some domains, something’s going to come up. Well, you can do it on DNA, you can do it on genomics. Or you could do it for video games. Yeah, actually, there are other domains that I didn’t mention, but you have to do the literature search. It goes back to the PhD training, what it trains you is to do the literature search.

So step one, find something you’re interested in. Sometimes it’s lost, but it’s so critical, it’s like, what are you interested in solving? Start to look up what people have done in that field, what the challenges are today, is it even an ML problem? Potentially, let’s say it is an ML problem, is it a deep learning problem, have people applied deep learning to it? As you do that literature search, you go deeper and deeper into the actual research articles like on archive or whatever, reading a bunch of blogs and things.

Then you start to craft a problem, but you’re passionate about the domain or the big overarching problem. And now you’re crafting maybe a problem that deep learning can be applied to, because it’s ultimately not magic. We’re not creating sentient beings, you’re-

Dave Cole

Not yet anyway.

Eitan Anzenberg

… solving a particular… Yeah. It will take some time, maybe Tesla’s trying, but it will take some time. It’s solving a specific problem, but it can, in some of these instances, can solve it really well. And then, because if you’re interested in neural networks you’ll probably start to dig into all the different types of layers. I mentioned convolutional nets, recurrent nets, transformers, capsules, there’s all sorts of many different ones in the literature.

So if you’re interested in these layers and you have a problem that you’re interested in solving, what is a good one? Let’s say you have a new baby, you have a baby monitor, and you’re trying to analyze that video and make it classify. Oh, my baby’s up, she stood up, so maybe she’s ready… and you want to send a notification. That’s a cool problem. You have something that you might be passionate about, get that video data, read the literature, figure out what kinds of models people have done.

Whether you run it in the cloud or if you have hardware on your desk, that doesn’t really matter too much. And then just start to experiment and see what you get. But it’s solving that problem. We don’t have a robot nanny. It’s like it’s solving a particular problem. So I think you always have to frame these ML problems in that regard.

Dave Cole

100%. I think that’s a recurring theme in the Data Science Leaders podcast, is to know the business problem, start there before you just start throwing code at it, throwing hardware at it, etc. Well, cool. Let’s wrap up one final topic. I mean, we talked briefly about the fact that you have a physics background, and you clearly have been extremely successful in your career despite getting your PhD in physics.

Eitan Anzenberg

Despite.

Dave Cole

Despite. Yeah.

Eitan Anzenberg

Exactly.

Dave Cole

Yeah. But, I’m curious, if you’re struggling to hire data scientists, one fruitful ground of folks out there, there are some extremely smart people who are getting their PhDs in technical fields and even psychology, you name it, that have statistics baked into them, and it behooves you to take a chance on some of these folks. Clearly, somebody took a chance on you, and it’s been paid off in spades, so I assume you would agree.

Eitan Anzenberg

I agree. I agree completely. One thing I like about being in this area is even just like the degree matters less. It does, it matters less than maybe other areas in the world like being a lawyer or something. I’ve noticed that it’s not the degree or where you went to school, and I like that. I actually like that a lot, because it matters more about how you solve problems or how you look at a problem, how you try to solve it.

There are a lot of untapped resources in the academic field, but also I have friends that are professors now or researchers or scientists in basic science. Not everyone wants to go and work for companies, not everyone wants to move away from academia, so I totally understand that. But, yeah, I would say if you have an opportunity or, I don’t know, how would you say so, if you have a connection with labs, I mean, we have a connection with Rice University and also with SF State. So if you have a connection and you start to talk with students, whether they’re undergrads or grads, and see what their interests are, are they interested in, for our group, machine learning, but maybe they’re interested in computer science, maybe they’re interested in mobile development, or maybe they’re interested in a whole other different thing.

The search for strong candidates, you never know where they’re going to come from, I guess. If you’re searching you have to really search everywhere. You have to go really deep and search because there could be strong candidates anywhere, even like, obviously internationally, with the pandemic people are remotely hiring everywhere. And so, yeah, so I wouldn’t limit, obviously, don’t limit to one particular undergrad or major.

Dave Cole

Yeah. I think what I’ve found in working with PhD grads is due to their role, they’re always learning, they’re loving to learn. One thing we know about data scientists and the data science community is that learning is part of the job, because I think the pace of innovation is so quick and it is so easy these days with open source and now the cloud to be able to experiment.

Experimentation has to be in your blood, and most folks who come from getting their PhD, that’s what they have been doing for many years. To me it’s fertile ground, it makes a lot of sense, take a chance, you’re not always going to be able to hire somebody with a PhD in stats, orI don’t even know if you need a PhD in data science these days, but who knows, maybe you can.

Eitan Anzenberg

They’re starting now. But that’s a really good point. That’s a good point, because in a PhD environment you’re tackling one problem for multiple years, and so you’re going to hit your head against the wall. I know I did for a year, two years, trying to understand that problem. And so you do build that. It becomes a muscle of experimentation and the research side of things is critical.

In machine learning, and especially in deep learning now, there’s a lot of trial and error. And even when you read articles you’ll see they sometimes contradict each other, like one article finds a particular architecture really good and another one doesn’t, and then another one actually affirms the first one. And it’s like a ping pong, back and forth. In other words, no one really knows what’s going on.

And then in academia, if you’re a grad student you’re producing literature that’s, in some cases, on the cutting edge of your domain of what you’re trying to study. So there is no blueprint, there is no right or wrong answer, and so you’re trying to figure that out. You do build that as a muscle.

And I have seen PhD people that went into the machine learning world, that they do have that capability and they are successful because of that, because they don’t give up. Maybe that’s cliché or whatever. When something doesn’t work they just keep going, they keep at it.

Dave Cole

Well, hey, Eitan, this has been great. I’ve learned a lot, I hope our audience has as well. I really appreciate you taking the time. If people want to reach out to you I assume they can reach out to you via LinkedIn. Is that correct?

Eitan Anzenberg

For sure. Yeah. Find me on LinkedIn. Send me a message if you’re curious about something or you want to chat, have an opinion.

Dave Cole

Awesome. Well, Eitan, thank you so much for joining the Data Science Leaders podcast, and have a great rest of your week. Take care everyone.

Eitan Anzenberg

Great. Thanks, Dave.