Following on from my conversation with Megan Butler about the state of the market for AI in recruitment, for the second part of this mini-series explaining AI and Machine learning, we are taking a deep dive into the technology and terminology itself.
My guest is Jon Krohn, Chief Data Scientist at recruiting technology company Untapt. Jon is also a lecturer, researcher and author in the field of machine learning, deep learning and natural language processing. This interview goes into a huge amount of detail into the definitions and mechanics of AI (in an understandable way!) as well as exploring current and future applications for recruiting.
In the interview, we discuss:
- The role of a data scientist
- AI and sorting the mail
- Building predictive data models
- Machine learning versus deep learning and why Siri got clever
- Data engineering
- Adoption and bias
- Current and future applications of AI in recruiting
Subscribe to this podcast in Apple Podcasts
Transcript:
Matt Alder [00:00:00]:
Support for this podcast is provided by Smart Recruiters, the hiring success company.
Matt Alder [00:00:06]:
Smart Recruiters offers enterprise grade recruiting software designed for hiring success. Move beyond applicant tracking with a modern platform that provides everything you need to attract, select and hire the best talent. From candidate relationship management to programmatic job.
Matt Alder [00:00:24]:
Advertising, recruitment, marketing, collaborative hiring and embedded artificial intelligence experience. A talent acquisition suite with intuitive user experience that candidates, hiring managers and recruiters all love.
Matt Alder [00:00:40]:
Leading brands like Bosch, IKEA, LinkedIn and.
Matt Alder [00:00:43]:
Visa use Smart Recruiters to future proof talent acquisition and expand their businesses globally. Visit smartrecruiters.com to find out how you can achieve hiring success as well.
Matt Alder [00:01:14]:
Hi everyone, this is Matt Alder. Welcome to episode 208 of the Recruiting Future podcast and the second part of my mini series explaining AI and machine learning. Following on from my conversation with Megan Butler about the state of the market for AI in recruiting, in this show we’re taking a very deep dive into the technology itself. My guest is Jon Krohn, who as well as being chief data scientist at recruiting technology company Untapt, is also a lecturer, researcher and author in the field of machine learning, deep learning and natural language processing. This interview goes into a lot of detail, into the definitions and mechanics of AI, as well as exploring current and future applications. And I’m very grateful to Jon for sharing his expertise.
Matt Alder [00:02:06]:
Hi Jon, and welcome to the podcast.
Jon Krohn [00:02:09]:
Hi there Matt. It is great to be on with you.
Matt Alder [00:02:11]:
An absolute pleasure to have you on the show. Could you just introduce yourself and tell everyone what you do?
Jon Krohn [00:02:17]:
Of course. So I’m the chief data scientist at a machine learning company called Untapt unt. Untapt, and we build systems that automate aspects of talent processes. So my team, the data science team here, we devise algorithms that allow talent processes. So these could be external recruitment pipelines, they could be internal recruitment initiatives, they could be identifying talent. There’s a large number of processes that today, with the kinds of scales of data that we have access to as companies or as recruitment agencies, we need models to be able to make the most of all of these data. That’s what my team and I develop. And then at Untapt, we also have an engineering side of our firm that then brings those models that automate these talent processes into production so that they either work as easy to use click and point screens, you know, usually administered via the web, or also we can have kind of custom, what we call APIs that which are kind of custom endpoints that then a client can, say, throw a candidate’s profile at it, as well as a job description and get back a score from 0 to 100 representing how good of a fit that candidate is for that job. So, yeah, in a nutshell, that’s what we do here at Untapt. And kind of the piece that I’m responsible for is managing the development of.
Matt Alder [00:04:08]:
Those models in our industry at the moment. There is obviously kind of a huge shift in technology going on as data science and machine learning and artificial intelligence and deep learning and natural language processing and lots of other bits of jargon and terminology that I could throw in as sort of being thrown around very liberally when it comes to, you know, some of the products are out that are out there and some of the solutions that people put forward. And I don’t really feel that as an industry and as buyers and consumers of this technology, we really actually genuinely understand what, what’s going on and what some of these, you know, what some of these phrases kind of actually mean. So, you know, I thought having a kind of a proper data scientist on the show, it would be great to explore some of this to, to really help people move people’s understanding forward. So first question, and quite, you know, quite a simple sounding question, and in some ways you’ve given us a bit of an answer already, but. But what does a data scientist actually do?
Jon Krohn [00:05:10]:
Great question. So data scientists, unlike say, physicians, don’t have a specific licensing exam in order to be qualified as a data scientist. So there’s a broad range of tasks that data scientists can do on the relatively simple end of things are what we call data analytics. So data analytics is familiar to most people. So even if you use Excel or some other spreadsheet tool for summarizing columns and creating bar charts, that’s data analytics. So data analytics is looking at historical data and summarizing it in some way so that you can visually identify trends that help you make better decisions. So data analytics is kind of the simplest part of being a data scientist. Another part of being a data scientist and the part that is kind of higher value and tends to be more complicated. And the part that’s hard to find qualified people for, often you need to have a PhD in a quantitative discipline to be able to do this proficiently is to build statistical models or machine learning models that are able to take in data. And where. Whereas the data analytics practitioner that I just mentioned is responsible for taking just historical data and making sense of those Data, a data scientist with, you know, these kind of predictive models that they’re building, these predictive models need, they’re being designed to predict events that have not been seen yet. So there are all kinds of approaches for doing this kind of prediction. So there are things like regression modeling, there’s things called decision trees, including a particular implementation called a random forest decision tree. There’s machine learning, and I’m going to talk about machine learning a lot, probably over the course of our time today, so I won’t get into that too much right now. But so there’s these various approaches for taking historical data that we can then use to predict future data. So that kind of probabilistic modeling is one of the core elements of data science. And that’s probably the kind of canonical way of describing what data science is. So taking in historical data, building models that can make a prediction of what might happen in the future given similar data. And there’s a, there’s a quote that I often talk about, which is by a famous 20th century statistician named George E. P. Box, and he said all models are wrong, but some of them are useful. So that’s kind of what happens here. So when you have a model of the world, when you build these predictive models, they are never going to be perfect. They’re never going to have 100% accuracy. Particularly in a field like recruitment or anything related to talent. These are very complex decisions where, you know, if you’re thinking about a hiring manager and them having a job description in front of them that they’re trying to fill and they’re considering 10 different resumes for that job description, the person, the people that they think are qualified or not. To some extent there is a clear rationale, but there’s also a lot of fuzziness where, you know, to some extent the number of people that they think are great is probably related to how recently they had a coffee. So it’s, it’s a, it’s not like, you know, these, these decisions that we’re trying to emulate with models in, in talent or in recruitment, these are not like physical processes like predicting, you know, how a photon of light is going to bounce off some surface where every single time you run the experiment, it’s going to go exactly the same way with anything related to human decisions. Where we’re modeling human decisions like this, there is a lot of nuance, a lot of complexity. But despite all of that complexity, provided we have enough data, we can build models that are more than accurate enough to be very Useful and to make hiring managers, recruitment agents, lives much simpler by cutting down repetitive tasks. And so I realize I’ve just spoken for a very long time, Matt, but I just want to very quickly say that. So now I’ve kind of described data science, the data science part of being a data scientist, but there’s another final part which is also critical, and this is what we call data engineering. So once you have, once a data scientist has built their model that they think works well, and we use programming languages to do that, the most popular one today is Python. And so we will use Python to maybe as a team, work together to look at the data that we have and consider different kinds of approaches for modeling those data. There might be several stages in a row that require different kinds of complex models in sequence in order to fulfill a particular client’s need. After we’ve done all of that work, we then need to do data engineering to bring that data science model to life. A model is only good if it’s useful and if it runs quickly. Data engineering is all about getting those models running in production systems, understanding cloud compute servers, understanding how to make a website, be able to call these models and work in an efficient way.
Matt Alder [00:11:09]:
And so, yeah, that’s fantastic. That’s certainly moved my understanding forward, and I hope it has for everyone else who’s listening as well. I want to kind of get more into that sort of human decisions and that, and that recruitment piece in a second. But before we do, just sort of coming back to your point about machine learning, machine learning, AI algorithms, these are all phrases that are kind of thrown around by salespeople a lot in our, in our sector. What is it important for HR and recruiting professionals to understand? How should they sort of be educating themselves about these particular terms? Is it important to know what the difference is? What’s your view on the sort of level of knowledge that we should have about some of, some of the jargon that’s, that’s kind of thrown around?
Jon Krohn [00:11:56]:
That’s a great question, Matt. And so much of it is jargon. We see so many companies that are AI companies these days. It’s definitely a buzzword that helps you get venture capital funding in the last few years. So we’re seeing that thrown around everywhere. So it’s my pleasure to be able to dispel the complexities around these various terms. So I’ll start with algorithm, because that’s kind of, that’s probably the easiest to explain. So an algorithm is any kind of process. So algorithms aren’t just confined to machines. Actually, so when you want to start your car, there is an algorithm for doing that. You need to sit on the, in the seat of the car, you need to press your foot on the brake, you need to find your keys, you need to put the keys into the ignition and then turn the keys. So that’s the algorithm for starting your car. So an algorithm is just a set of rules for carrying out some task and it ends up being, the term ends up being used a lot in anything related to computers because computers are excellent at being fed a particular sequence of orders, an algorithm, and executing on that. So that’s what an algorithm is. Artificial intelligence is probably the broadest of all of the terms that you mentioned there. So artificial intelligence is a buzzword for sure. It has a fuzzy definition and that definition is actually always changing. So in the 1990s, things like recognizing handwritten characters. So up until the 1990s, all mail had to be sorted by hand. But in the 1990s, we came up with algorithms with that were able to recognize handwritten characters so that based on the postcode or the zip code on the envelope, the mail could be automatically sorted by machines. That’s what we call optical character recognition. In the 1990s, that capacity was considered to be artificial intelligence. But now, 20 years later, that kind of capability is so commonplace that it’s no longer considered to be AI. So AI tends to be something that describes the capabilities of machines that we’re just on the cusp of. So today that could be things like self driving cars, it could be machines that are very, very good at playing particular board games or video games. The machine translation algorithms, that’s what we think of as artificial intelligence today, or that’s how it’s used by the popular press. So yeah, so artificial intelligence is definitely a buzzword. But if we want to take a stab at defining it, it’s any kind of machine process that replicates some intellectual capacity of a human. And today we only have what we call artificial narrow intelligence. So artificial narrow intelligence is a machine being quite capable at some very narrow, specific task. So driving a car, playing backgammon, balancing a stock portfolio. If you were to take your, your, your algorithm for driving your car and try to have it balance your stock portfolio, well, it of course wouldn’t be able to do that at all because. So each one of these artificial intelligence capabilities that we have today in machines are very narrowly defined. There’s a lot of excitement that possibly in many of our lifetimes, something called artificial general intelligence will be attained, which is the Kind of AI that we tend to see in films. So this is like a robot, often a humanoid robot, that has all of the intellectual capacities of an individual person. So this is a single algorithm, a single model that would be able to drive you to work, converse with you, translate a document for you, and balance your stock portfolio and beat you at chess, all of these things at the same time. So it’s, you know, so this, this artificial general intelligence idea, it’s something that is only hypothetical. However, in a survey of AI experts, they speculate their, their median guess as to when that artificial general intelligence might be attained could be around. Well, their median estimate was 2040. So. But, you know, there’s all kinds of roadblo to attaining that kind of AI. There’s hundreds and hundreds of hurdles that have to be overcome to get there. And, yeah, so it’s only theoretical today.
Matt Alder [00:17:30]:
What about machine learning and deep learning?
Jon Krohn [00:17:32]:
So machine learning alongside some other fields like robotics, sorry, robotics. So machine learning in robotics are subfields of artificial intelligence. So machine learning is concerned with the software side of things, and robotics is concerned with the hardware side of things. So, you know, so robotics is, you know, actually designing a physical robot that can act in the real world and, you know, turn a doorknob and shake your hand and take care of an elderly person or so on. So that’s robotics. Machine learning is concerned with the software part of it. So it’s concerned with machine vision algorithms, so being able to recognize your face using camera sensors, natural language algorithms, so being able to understand the meaning of a resume or a job description. So machine learning is a field where we set up computer software in a way that when we feed it data to learn from, the machine learning algorithm is able to learn patterns in the data so that it can perform some task that we’d like it to perform. So I already gave this example a little bit earlier, and it makes sense for the purposes of, you know, your audience. So an example of a machine learning algorithm that we use here at Untapt is an algorithm that has two inputs. It takes in a job description, it takes in a candidate’s resume, and it considers the natural language, the words on those documents, and then it does some computation, and then it can provide an output, a probability as to the likelihood that that person will be invited to interview at that particular job. So in order to build this machine learning algorithm, which we’ve iterated on in various ways over the years, so our company’s five years old, I’ve been here for four years as the chief Data scientist. And from the time that I joined, my job has been to design models like these. And so the data scientist or the machine learning practitioner comes up with an idea in their mind, a framework, a model. So it could be. I mentioned the word regression earlier. This is kind of, that’s a statistical approach that many people have heard of. So, you know, so it’s just some kind of model. Some of them are very sophisticated. I’m going to talk about deep learning models in a moment. That’s another, that’s another example of a kind of machine learning model. And so the data scientist and machine learning practitioner comes up with this model and then they start feeding the data into it and they have their outcome that they’d like to predict. So in that case of the, the machine learning model that I just described, where we’re taking in the job description and the resume and outputting some probability that the person will be invited to interview at that job in order to build that model at untapped. We have hundreds of millions of training data points that we’ve accumulated over the years of particular people with particular resumes being invited to interview or being not invited to interview for particular jobs. So in each of those hundreds of millions of cases, we have a resume, we have a job description, and we have the outcome that we’d like to predict. We have a yes or a no as to whether that particular person was invited to the job or they weren’t invited to the job. So by training the machine learning model on hundreds of millions of cases like that, it can build up a really nuanced and complex understanding of the kinds of people that are an appropriate fit for a particular kind of job, and it can give a reasonable estimate of how likely that person is. So this probability, from 0% to 100% that any given person will be invited to any given job.
Matt Alder [00:22:02]:
And what’s the difference between machine learning and deep learning?
Jon Krohn [00:22:05]:
Yeah, so deep learning is a specific kind of machine learning. So I mentioned regression models. Now, another kind of machine learning model is what we call random forest, or there’s support vector machines. So those are examples of different kinds of machine learning models. So when you have your particular inputs and your particular output from your model that you’d like to predict, data scientists have, they spend years studying, okay, well, which of these kinds of approaches, Regression, random forest, support vector, machine. Which of these kinds of models are going to perform best with the data that I have and give me the most accurate outputs? So deep learning models are just another type of machine learning model alongside Those others that I just mentioned. And deep learning models are particularly exciting because since 2000, so deep learning models have been around for decades, and they’re based on the idea of these things called artificial neurons. So since the 1950s, we’ve had a reasonable understanding of how biological brain cells function. And since the 1950s, people. Well, specifically, somebody named Frank Rosenblatt came up with this idea of, I want to build a computer that simulates the way that brain cells work. And so he invented the first artificial neuron in the 1950s, which is just a really simple algorithm that is inspired by the way that biological brain cells work. And I don’t have time to go into the detail of that, but in the same way that our biological brains consist of billions of biological brain cells that are interconnected in the same way, you can take these artificial neurons, these artificial brain cells, and link them together into a network, into an artificial neural network. So if that artificial neural network has several layers of processing, so it’s easiest to kind of provide an analogy with respect to machine vision. So if you build a artificial neural network where you have a first layer of artificial neurons that are able to recognize very, very simple aspects of the world, like straight lines, then you can have a second layer of artificial neurons after that. That takes all those straight lines that the first layer identified and recombines them into more complex shapes, like curves and corners. Then you can have a third layer of artificial neurons that can take those curves and corners and make them into even more complex shapes, and a fourth layer that even more complex, even more abstract, until after several layers, you can have your artificial neural network. Take a guess that, oh, hey, that’s Matt’s face, or that’s Jon’s face, or that’s my grandmother’s face. So if you stack your artificial neurons in that way, such that you have at least five layers of processing, that is a deep learning network. So a deep learning network is a very, very specific. So whereas AI is a very vague term, and machine learning is a pretty specific term, deep learning is a very specific term that describes this very specific kind of building, a model where you have several layers of these artificial neuron algorithms. And the exciting thing about deep learning is that even though it’s been happening since the 1950s, by 2012, the cost of computing had gone down enough, and the cost of storing large amounts of data had gone low enough, and a few theoretical advances had happened. So the confluence of those three factors, cheaper computing power, larger data sets, and a few theoretical advances enabled in the year 2012, deep learning algorithms to suddenly become the most potent way of building a machine learning model, of building an AI system. So the first big breakthroughs were in 2012 when deep learning models absolutely crushed all existing benchmarks on machine vision algorithm accuracy. So just the ability for a machine vision algorithm to be shown a photo and say, what’s in this photo? And answer that question accurately, it absolutely annihilated any existing approach at that time. And then by 2014, those advances started to make their way into natural language algorithms like Google Translate or Siri voice recognition on your phone. So all of a sudden deep learning started to show up everywhere in all kinds of applications, and it made these machines just so much more sophisticated and nuanced and correct. So if you think About Siri in 2010, 2011, Siri was not powered by deep learning on iPhones. And the number of mistakes that Siri made in made it so that you were like, eh, I don’t really feel like using this. But by 2014, 2015, because of the integration of deep learning models into the voice recognition capabilities that Siri on iPhones has, all of a sudden she rarely makes mistakes. And so it seems useful to be using this tool. So in that sense, the present revolution in artificial intelligence that’s happening is being led by deep learning, which is that very specific approach that I outlined of having layers of artificial neurons.
Matt Alder [00:28:00]:
I mean, that’s absolutely fascinating and it’s funny as well because I’d forgotten how bad Siri was and the kind of, the dramatic turnarounds, turnaround in Siri and kind of Alexa’s fortune in a really short space of time. And that now makes sense as to why that’s happened. I’ve got huge. I could literally sit and sort of ask you questions all day, but sort of, you know, with it, with, with an eye on time, I think I’ve sort of got a couple, a couple more. More things to sort of get your view on. It sounds to me that we’re barely scratching the surface of what might be currently possible when it comes to this technology and recruiting. So I’d be interested to sort of firstly get your view on that and then secondly, get your view on perhaps the sort of the. The negatives around this, including I suppose, bias and accentuating bad processes and things that don’t make sense. What’s your sort of take on those two, Those two things?
Jon Krohn [00:29:03]:
Great question. So yes, so we are definitely just beginning to scratch the surface of how models can be used to automate aspects of human resources or talent or Recruitment. So there are so the kinds of problems that are relatively tractable today, like that one that I’ve described. And it serves as a good use case because lots of different kinds of people are interested in it. So we have corporate clients that get millions of applications a year to thousands of different roles. We have recruitment agency clients that have thousands of recruiters working on who knows how many different kinds of roles and have databases of millions of candidates. And so those two things, corporates, recruitment agencies, as well as human resources technology companies. So HR tech companies that build screens, CRMs that enable you to manage your recruitment processes. So those three kinds of clients all have a huge need for models that can, for example, take in a job description and a resume and output a probability that a certain person is a good fit. Because for those different kinds of use cases, there are particular use cases that reduce a lot of headaches and a lot, and make processes run a lot more quickly for people in those companies. So in the case of the big corporation getting millions of applications, they can, a hiring manager can show up at her desk on Monday morning and she can find out, okay, 1,000 people applied over the weekend to the five roles that you have open. And now instead of her having to look through all of those resumes in a kind of random order, maybe in the order that they applied, you can use an algorithm like ours to sort by the ones that are the best fits so that you can start at the top of the list. And of course, people who are a great fit, they’re likely to be hired by somewhere else quickly. So you want to be able to make those moves quickly. And see, okay, start from the top of the list. The people who have a 99% chance of being a good fit will start with them. So, yeah, so that’s kind of a corporate use case with the recruitment agencies. You know, a similar kind of thing, sorting resumes, allowing parts of their processing to happen automatically. So one client of ours, they have all of their, all of their recruitment consultants fit into particular categories that they specialize in. And so we devised a model for them that whenever somebody, you know, shows, expresses an interest in working with their recruitment agency, that resume automatically gets categorized as being in a particular category like finance or pharmaceuticals or what have you. And then the right recruitment consultants are made aware of that new individual. And then the kind of final example, there are the HR tech companies who are building screens for managing recruitment processes. They can take advantage of these kinds of algorithms so that their clients are able to see, oh, hey, for our software Engineering pipeline of applicants, we are getting a large number of high quality resumes. But for our, for our accountants pipeline, although we’re getting a lot of applications from accountants, most of these applications are low quality. So we need to ramp up our efforts on marketing to accounting fairs or whatever to get higher quality applicants. So that kind of data pipeline monitoring. Anyway, so those are some examples of the ways that these algorithms can be used to kind of bring these ideas to life. But in any of those use cases, as you mentioned, there is of course a risk of bias. So you know, I mentioned to you, some of our models are trained on hundreds of millions of data points and those decisions to interview or not interview somebody for a given role that are used as the training data points for our models, there could certainly be biases in there, right, where you know, a hiring manager has a gender bias or a racial bias or what have you. And so a huge part of what we have to do at Untapt, and I suspect any kind of company that is building models that involve modeling human decision processes, we spend a huge amount of our time devising these models to work in such a way that they do not replicate the biases that are inherent in human decision making. We actually just two weeks ago we filed a patent in the US to cover our particular process for removing bias from predicting candidates suitability for a given job. And we spent four years really devising that approach. So it’s a complicated thing to get right, but it’s really rewarding when we can now run tests and say, is there a higher probability of a female applicant or a male applicant getting a high score in our model for a given role? And it’s really rewarding to be able to say no, absolutely not in that way. If these models are built in a way that takes into account these biases in the, in the data that the algorithm was trained on. And if you work to ensure that those biases aren’t present in the outputs, you actually have the capability with machine learning models to remove biases that are otherwise difficult to track. So you know, up until recently, you know, we weren’t tracking data on are we interviewing as many female applicants that apply? So the proportion of women that apply to this role, are we interviewing that same rate or are we interviewing fewer of them? Those kind of data, that’s something that’s only recently something that people are keeping track of. These models, if they’re built appropriately, these machine learning models enable you to go one step further and not only monitor whether those biases are there and Then kind of retrospectively take action. They in real time eliminate the biases and enable us to have less biased recruiting.
Matt Alder [00:36:25]:
Final question, and again, might be quite a difficult one to answer, but where’s this going? If we were sort of fast forwarding five years, what might we be able to expect that we could achieve with these types of technology in recruiting?
Jon Krohn [00:36:41]:
That is a great question. So there are a lot of ways that this can go. So the big thing over the next five years is going to be adoption, Matt, because it’s super interesting how many clients we speak to that, or I should say prospective clients we speak to that are aware that they need an algorithm like the kinds that I’ve described in order to automate something that they’re doing that’s an incredibly tedious, expensive manual task and an inaccurate task to do manually because of fatigue from, say, screening tons and tons of different resumes. And they also know that there’s this problem of bias in that kind of resume screening process. So they’re aware that that’s a problem. But the vast majority of prospective clients that we speak to, they don’t have systems in place where they can even adopt an algorithm today. So they don’t keep track of the data that they, that they have in a way that enables them to even take advantage of a model. So they might have the resumes just strewn across various directories, on various hiring managers, computers across their corporation. And so the first step for any of those companies is just building a database that is consistent and automated across their company. And so that’s the first step. And it’s amazing that in a lot of companies is where they’re at just creating these databases. So that’s the first step, and we can help with that kind of thing. There’s all kinds of companies out there that can help you make a database of your human resources data. And then step two is then starting to build these modeling processes on top of that that enable you to automate parts of these processes, parts of these human resources tasks, so that adoption, it is never straightforward. So every client, they store their data in different ways and they have some different outcome that they’d like. They have some different internal initiative that requires that internal initiative to be handled in some way and, and made clear on their user interface screens in some way. So even the way that companies, and this is particularly large companies want to adopt these automated processes. There’s so much kind of bespoke needs in all of those cases, some of which are regulatory, some of which are simply preferential, that over the next five years, the vast majority of the change is just going to be adoption. So getting people’s databases sorted out properly and being able to layer some models, some simple, relatively simple models like the ones I’ve described during our time today, on top of that, that’s going to be the vast majority of the next five years. However, beyond that, in terms of looking towards the future, there is going to be. So every 18 months the amount of data that we have on earth doubles. So if you think about the amount of data that we have today, 18 months from now, we’re going to have double that. It’s staggering. So self driving cars generate a huge amount of data and your phone generates a huge amount of data. And so more and more we have these devices that are generating more and more data people today in terms of recruitment. They might do video interviews that are recorded, that generates a huge amount of data. And they might do tests that generate data. And fields generate a huge amount of data. So we have, over the next five years, in addition to adoption, we’re also going to have staggering amounts of data being generated in the world in general as well as in recruitment and human resources specifically. So integrating those data into a sensible outcome. So if you have, for a single candidate, you have the video recorded interview where you can be applying some kind of automated emotional analysis to that video. You might have their resume and the job description, which you can do some automated analysis of what kind of good fit the person is. You might have some tests that they filled out that gives you a sense of what a good fit that candidate is for your team in terms of personality or quantitative skills or something. So integrating all of those data sources where we have all these huge amounts of data bubbling out and having those result in actionable summary metrics and sensible automated processes that actually reduce headaches as opposed to increase headaches for hiring managers and recruitment consultants. That’s going to be another huge part of the next five years. And then just to kind of, to kind of say an exciting thing that’s beyond that is. So we have the set of. I just described a number of different kinds of data that are collected today on applicants for roles. But going forward, we’re going to have even more. This is something that’s inevitable. So with the rollout of 5G networks globally, the amount of devices with sensors is proliferating hugely. For example, my company, we are based in a WeWork space and WeWork has sensors in all of the common spaces that detect how often people are in those common spaces. And that gives them a sense of, you know, how much people are interaction interacting, how much they’re using the kitchen versus the kind of more, you know, living room style sitting spaces. And that enables them to optimize the layout of their buildings. Both simple things like the furniture, but also the way that they build. Future wework offices are informed by the data that they collect on all these sensors. So that’s an example of the kind of thing that going forward we’re going to have more and more of these kinds of data and even in ways that we can’t anticipate today that will enable companies to lay out their offices in a way manage their employees in ways that not only increase employees productivities but also increase their happiness in order to ensure that they can retain talent. So lots of big changes coming ahead over the coming decades. But yeah, because of the pace at which these things are changing, it’s difficult to know exactly how it’s going to look.
Matt Alder [00:44:21]:
Where can people find you and your work?
Jon Krohn [00:44:23]:
For all the work that I do with Untapt so all of this talent and human resources model development and building screens to allow those models to come to life and be useful for corporates, for recruitment agencies, for HR tech companies, all of that work I do as the chief data scientist at Untapt Untapt and so you can look that up@untapped.com or you can email me. I’m Jon Jon untapped.com on top of my kind of day job, I teach. So I teach engineers at Columbia University how to build these deep learning algorithms. I have my own deep learning curriculum that I devised and I teach on Saturdays at an academy here in New York called the New York City Data Science Academy. I do research at Columbia University into medical imaging and so automating aspects of medical imaging work. And I also write. So I have my first textbook is coming out in autumn 2019 and it’s being published by Pearson. And so for anything related to that kind of work, I maintain a website, johnkrone.com, j o n k r o h n.com so this first textbook is an introduction to a lot of the kinds of things that we talked about today. So it introduces what artificial intelligence is, what machine learning is, what deep learning is, and then it covers most specifically deep learning. So how to build deep learning algorithms for machine machine vision, for natural language processing, for playing games, for generating novel artworks. And I deliberately designed that book so that the first four chapters don’t require you to have any programming experience or any specific experience with mathematics. So if you’re interested in learning about what deep learning is in detail and what all the applications are that are possible today, check out that book, Deep Learning Illustrated. And yeah, I think that about wraps it up. Matt, thank you very much for these great questions. I hope your audience finds it really useful.
Matt Alder [00:46:42]:
My thanks to Jon Krohn. You can subscribe to this podcast in Apple Podcasts or via your podcasting app of choice. The show also has its own dedicated app, which you can find by searching for Recruiting Future in your App Store. If you’re a Spotify or Pandora user, you can also find the show there. You can find all the past episodes@www.rfpodcast.com on that site, you can subscribe to the mailing list and find out more about Working with me. Thanks very much for listening. I’ll be back next week and I hope you’ll join me.
