00:00:10 Mae Macfarlane
Hello. I'm Mae Macfarlane, here with the professor podcast, the podcast of your professors, their research, and their academic lives here at Saint Thomas
This week we are delighted to have with us Dr. Manjeet Rege, a professor in the Saint Thomas Software and Data Science Department.
Welcome Manjeet.
00:00:26 Manjeet Rege
Thank you. Mae, it's my pleasure to be here.
00:00:29 Mae Macfarlane
Very excited to talk with you. Can you tell me a little bit about your journey, about your background, and where you got to where you are today.
00:00:39 Manjeet Rege
Sure! So, I have a very diverse academic background. Have an undergraduate in mathematics.
From there I at the end of my master, end of my undergrad in mathematics, I got some exposure to computing, and I figured that probably I should learn a little bit more. So, during my masters, I decided to pursue a Masters in Information Systems, so it's more like a business degree.
So, there I got to know a lot about computing and at the end of my masters I thought, you know, this is really interesting. Can I get a little bit more technical? Perhaps an advanced degree.
And I decided to pursue a PhD in computer science. Computer science itself is a very large field, so within that I initially started with database and those days, this was 20 years ago, there was relation database, so when you talk about the relation database, think about tables, for example, and then there was something called XML databases.
And you know as a PhD student you're assigned to the project to work on, so I was assigned to work on a medical image database for neurosurgeons. So as part of that, you know, I had to attend a neurosurgery for many hours, just to watch what that process is. Because in order to really do something, you have to...It begins with understanding the process. So, I...it was an exciting thing to attend neurosurgery and I was working on the medical image database.
00:02:26 Manjeet Rege
And as part of that, you know, once you create a database you have to then think about deriving value from it.
How do you then query information? So that is where other techniques like machine learning come into the picture. And so eventually my dissertation, PhD dissertation topic was in machine learning and that is how I got started in machine learning.
Now I teach courses in data analytics and visualization, artificial intelligence, machine learning
00:02:51 Mae Macfarlane
Can you explain a little bit what machine learning is?
00:02:56 Manjeet Rege
Sure. Let me start with AI first and machine learning is a part of it.
AI is a broader field, kind of like the field of physics.
00:03:07 Mae Macfarlane
And AI is artificial intelligence.
00:03:09 Manjeet Rege
Artificial intelligence.
So, typically you know, we use three abbreviations AI, ML, DL.
So artificial intelligence, machine learning, deep learning.
So, AI is the broader field. As I said, kind of like the field of physics. And the goal of artificial intelligence is to create computers or computer models that behave or mimic like humans, not necessarily to replace humans.
So, in order to perform automation at scale. Now the field started in around 1950s. Now within AI there are many subfields. One such sub field of AI is machine learning that deals with learning from data. So, machine learning essentially means if you provide me lots of examples, then a computer model can learn the associations.
00:04:10 Manjeet Rege
If you give me one picture of a car. And then you have a label next to that, saying that this is a car. Another picture of, let's say, a table and there is a label there saying that this is a table. And you have thousands and thousands of such examples with the image and what its label is. A machine learning model can learn that association -- that whenever the shape or the color or other features appearing are in a particular way, the object that is appearing in that is a flower or a car or a table. So, that is what machine learning essentially does. It learns from data.
Now, there are a number of machine learning techniques out there, such as logistic regression, support vector machines. There is something called neural networks. So, neural networks is one such technique that has been inspired from how our human brain is structured in terms of neurons.
00:05:12 Manjeet Rege
So, if you think about the hierarchy here, AI is a broader field; machine learning is a sub field. Machine learning consists of a number of techniques. Neural network is one such technique and when you make a neural network large in the sense of its compute power and complexity, it's called a deep neural network, which in modern days we call it deep learning.
So, you have artificial intelligence. You have machine learning. And you have deep learning: AI, ML, DL.
00:05:47 Mae Macfarlane
That's really interesting. I didn't. I always kind of assumed they were separate.
And now that you say that they're, you know, they gradually go down, that makes a lot more sense.
I have a quick question about that.
So I don't know how much you're on Twitter. There is a user who people could submit images to and the AI, well, I guess it would be. It's an AI that's using machine learning to figure out what the image is.
One that I'm thinking of is a picture of a bunch of ping pong balls in a fridge and the AI guessed that it was eggs at a 98% guess? It was very confident. How does that wide data set of people sending things in randomly, without assigning a name next to it, what does that show us about the machine learning?
00:06:30 Manjeet Rege
So what you're talking about is people are tweeting images to this particular Twitter handle, and so that that Twitter handle is passing that image on to a particular AI model, getting the label that the model is coming out with and then tweeting it back as a response.
00:06:50 Manjeet Rege
So that is a model performing prediction. It typically the model initially goes through a learning phase. And then there is a phase where, and you know that the model has learned, it can now make a prediction.
So, think about how a human or a child learns. Initially the child, let's say, has never seen how a car looks or how a table looks. So, you show a lot of examples of different cars and you, every time you see a new car you just point. And say “that is a car;” ”that is a car.” “This is a table.” And every time you are showing different kinds of cars, different kinds of tables.
00:07:29 Manjeet Rege
So, over a period of time when the child sees a lot of examples, the child is able to differentiate between these two distinct real-world categories, also able to appreciate that there are different kinds of tables, some of them could be rectangular, some of them could be wooden, glass.
So now that the child is learned, the model training is done and now if you give the child a new image of a car, the child is in a position to give you that prediction and say that's a car, or that's a table, or that's something else. That is the use case that you just now specified.
00:08:08 Manjeet Rege
Now sometimes what happens is we make mistakes, right, in in our prediction. A model also does mistakes, and even though the model says that there is a 98% accuracy, sometimes it gets it wrong.
So, when you have a model in production and you notice that it is misclassifying and classifying something as a particular object when it is not, then you need to go back and retrain that model.
00:08:40 Manjeet Rege
You need to first of all collect all of those images that the model is making errors on, and then probably you might notice that all these images that it is making errors on would click using a particular let's say Instagram filter. And the model during its training phase has never seen images with an Instagram filter and that is why it is giving you a wrong prediction.
00:09:08 Manjeet Rege
So the way to fix that is to retrain, get a lot of images, now with an Instagram filter as well. And now when you deploy it, you might notice that it will get the right predictions.
00:09:22 Mae Macfarlane
And thank you for explaining that. I know that was kind of a goofy real-life example, but there's a lot of accounts like that right now where people, like college students with computer science degrees are like, let's take it out of R and put it on Twitter and see what we get.
00:09:38 Mae Macfarlane
Speaking of like being a student, learning about all this, how does one learn to become an AI practitioner?
00:09:45 Manjeet Rege
So an AI practitioner is not that of a long journey that one might initially think about. Things have changed rapidly in the past 10 to 12 years. If you look back maybe 15 years ago, in order to become an AI practitioner, you had to go through a long, rigorous curriculum. AI wasn't accessible to high school students, for example.
But now, with cloud computing, you have so much of compute power at your disposal, you can go to any of the cloud providers like Microsoft or Google.
00:10:28 Manjeet Rege
Also now in order to build your own AI model, you don't have to write a lot of code. That was not the case 15 years ago. So, you had to write lots of code and many times the code that you write could have errors in it. So, it wasn't something that you could put into production in the real-world, just like any technology that develops, it becomes more and more accessible.
So now with cloud computers taken care of, with the Internet of Things, and as well as you know all of us being data generators with the phones that we carry, more and more amount of data that is available, it helps an AI model to learn.
And with these two restrictions being taken away, that these restrictions were there in the past and as a result, a lot of AI/algorithmic research could not advance, because if you think about a new AI algorithm, you were restricted by compute power.
00:11:33 Manjeet Rege
You didn't have much data, so because those two doors opened, there were a lot of advancements that happened in the field of AI, and there were a number of libraries that were developed. So when I think about libraries, think about the development of AI model became accessible to software engineers or even you're going towards a low code or a no code AI developed.
00:11:58 Manjeet Rege
So now it is possible to have just about enough knowledge about AI and being able to develop a model.
So, it really depends on how deep you want to go in your AI journey> Within our curriculum we have a graduate certificate in AI in my department. It's a 5 course sequence. We start with actually 4 core sequence.
Now from the fall you just need to know Python programming language. And some knowledge about statistics, which is part of that certificate, machine learning and AI, and you could easily become an AI practitioner if you want to develop your own AI models using code.
But let's say you do not want to and you want to you are, let's say, a business student and just want to build AI models directly. There are a number of tools and platforms that allow you to build AI with low code or no code. As well, tools like let's say Aletrix, there is something called Data IQ. Google Teachable Machine is another resource online that allows you to select a number of images that you want. Just upload it directly and those images get the – the model learns to classify those images directly.
00:13:23 Manjeet Rege
So there is something called auto ML or auto machine learning which is where the trend in recent times has gone to where the bar has now kind of gone down significantly in terms of making an entry into the field of machine learning or AI and also being successful at it.
But if I have to simplify things further, knowing certain things like let's say a Python programming language helps you quite a lot. If I have to pick a language of let's say data science or machine learning or AI, that would be python so knowing that….And I know number of even high school kids now know Python very well…will give you head start, but if you don't know that, that is not end of the world either, you could still do some no code AI as well.
00:14:11 Mae Macfarlane
So, you mentioned like business students wanting to get into this or I mean, I feel like it could be a very intersectional piece of technology for many people to use.
What should people do to learn more about it?
Say they've graduated and already in the workforce and they think, “Oh, my business could use this? My team could use this.” What should they do to pursue that?
00:14:32 Manjeet Rege
So I'm going to put on my hat of an educator and see that a college education, you know, prepares you on the whole as a person, right? It is not necessarily preparing you for, you know, lifetime job employment. It prepares you as a whole. It prepares you for being a good learner as well.
00:14:55 Manjeet Rege
So, one of the things that I always say is being a lifelong learner, is no more an option. We all need to be lifelong learners. So, many students in our department, so we are a graduate only program and we cater to career changers. Students who have had an undergraduate in, let's say, literature, history, biology, business and they want to pursue a degree in data science or AI. Go through our curriculum, and once you go through that curriculum, yes, you become an expert in that field.
But remember that that field is constantly changing, as well. So, keeping yourself abreast with technology is the key. At the same time, taking something which is going to advance you professionally on a regular basis.
It could be attending professional development conferences, trainings, workshops, seminars, going back to grad school. And I understand that for some it might be overwhelming to go back to school and commit themselves to an entire masters deck.
But it could be even one or two courses that can make a significant difference. It could be a small graduate certificate that one can start out with as well.
00:16:15 Mae Macfarlane
What is the responsibilities of universities as a whole, not just Saint Thomas in this learning? Like what can schools do?
Is there a certificate programs or a masters plus kind of program that you kind of can learn? What do you build the responsibility?
00:16:29 Manjeet Rege
Right, right.
So a couple of responsibilities. One is to help people embrace this change, when you look at a lot of companies, they have their workforce there. They have a lot of domain knowledge about their business, but they may not have an AI skill set.
So catering to people like those which our department has been doing in terms of helping them transition into another career. Also, when you look in the media, AI also sometimes gets a bad rep with deep fakes and things like that.
You have so much access to data and you have such stronger AI models that as educators we are also always telling students that when you get access to large amounts of data and you're building strong AI models.
It is important to keep ethics in mind as well, and AI and ethics is extremely important. How do you make sure that a model does not have, or has minimal bias?
When you show more examples of a particular category and less of the other category, the model automatically develops a bias towards one of those and you know you will never get it correct in the first go. But your first iteration of the model, if it is working that is great, but you have to constantly refine the model and make sure that it is not developing a bias.
So that is the second responsibility I would say as an educator.
00:18:11 Mae Macfarlane
And then how does it impact your teaching?
00:18:14 Manjeet Rege
So it is a part of what we do. In as far as AI and ethics is concerned, in my AI course I talk about examples where AI models have developed a bias and I talk about how do you make sure that could be minimized?
There are many real-world examples of AI model bias, so when you bring those examples into the classroom, students are actually able to relate to problems that a model bias can create in the real-world.
00:18:47 Mae Macfarlane
You've spoken on a lot of different media platforms through interviews with local Twin Cities news stations and such. When you're doing those interviews and when you're doing this interview, what do you want the public to get from what you're saying and what you're teaching and what you're researching?
00:19:04 Manjeet Rege
So what I want people to notice when I've I've spoken on a number of issues, whether it is the disparity by race or data visualization for example When I typically speak in the media, I only speak from the data perspective and I do not draw conclusions based on that, one way or the other, that is not my position or my expertise.
But many times you see the limitations of analyses. For example, if I take the example of Apple’s credit card and Apple had a credit card that used to give a higher credit limit to the husband, but less credit limit to the wife.
Now that is an unfortunate case of a model developing bias. It wasn't somebody's intent, probably, but just that they had more examples -- Remember, an AI is learning from data, so they had more examples of males, less of females or males perhaps making more. and the model just learns from that and it says that “OK, this is what you're giving me then probably that is how I should be predicting’
What I also want people to take notice of is that when a model perhaps gets it wrong, humans make errors. So many times when you have a computer making a prediction, people expect a computer to be 100% accurate because it is a computer.
But you know, if a human makes an error, we Say “Yeah, that's all right. You know, we all make mistakes,” but a computer is not, or an AI model is not as strong as a human brain is.
00:20:52 Manjeet Rege
So it makes errors and over a period of time it learns. So we have to be kind of forgiving and an AI model as well. As long as humans we are able to go back and fix the problem.
00:21:04 Mae Macfarlane
So, I'm 21. I've grown up kind of in the age of technology, mostly in the age of technology. And you know, we hear a lot of that, like AI is taking over jobs, blah blah doing all this stuff.
But I think what you just said is important -- that while it can do all these great things, it's still an imperfect system.
What is the impact of AI on jobs since it Is not perfect and still needs the human brain to kind of keep it pushing along?
00:21:29 Manjeet Rege
One of the misconceptions that people have about AI is that AI is out to replace humans. And so we are trying to create artificial intelligence that will replace human intelligence and that is not the case. In spite of the advancement that we have seen in the field of AI over the past 12-13 years or so. Which has been significant compared to, you know, when the field started. So last 12 years have been extremely productive.
In spite of that, if you look at what a human can do and what an AI can do, there is a big difference. A human intelligence is what we call a generalized intelligence and where the AI field has advanced to is what we call narrow intelligence.
When you look at your phone and the phone unlocks itself, or when you sit in a self-driving car, drives, navigating through traffic and stop signs. Or, you know, many other applications. All these are different AI models. It is not one AI model.
00:22:42 Manjeet Rege
If you think about a human. That one human can do all of these things.
00:22:47 Manjeet Rege
Also, when you think about learning an AI model takes a lot of time, a lot of examples to learn. If you show a human, this is how a car looks and you bring variations of that category car. A human can quickly learn. “Oh yeah, I think this looks similar to what I saw.”
But yeah, AI needs a lot of examples to learn, It doesn't learn from a few examples. So that's what I mean by generalized intelligence and narrow intelligence.
So, we are currently at narrow intelligence, and we may or may not reach that stage of generalized intelligence.
00:23:29 Manjeet Rege
So, Speaking of jobs, if one has to replace a human, you need to have generalized intelligence, and only then we I mean, instead of you, there could be a robot sitting there instead of me.
And there could be a robot. And there could be 2 robots talking to each other: a professor of art and a student robot, right? But I don't think that's going to happen anytime soon.
Now, yes, there is an impact when you talk about jobs and I like to make the distinction between AI displacing jobs and AI impacting certain category of jobs, where those jobs you will get exposed to AI, so getting displaced by AI and getting exposed to AI are two different things.
So, there would be certain blue-collar jobs that may be at risk sooner of getting displacement. So, let's say jobs where there could be a lot of manual labor, in terms of, let's say, agriculture, long driving, let's say like truck driving, warehousing. In warehouses where you're heavy lifting, you know, you can have robots.
Amazon's warehouse is already doing.
00:24:40 Mae Macfarlane
That there's an example --This is very niche and like I said, kind of from my childhood.
But the Willy Wonka would be the failing the Chocolate Factory. His dad lost his job to robots, and that's kinda got ingrained into your childhood.
You know, this is a bad thing because your dad will lose his job.
So I think that that's really interesting. Kind of goes along those lines.
00:25:00 Manjeet Rege
So, the in that case, you know, the dad could have gone to Saint Thomas comma AI practitioner.
00:25:08 Mae Macfarlane
So I got you off your point.
00:25:09 Manjeet Rege
No, no, no, they're fine.
00:25:10 Mae Macfarlane
This is like what other impacts does it have? So like you're talking about like blue collar where?
Where it can easily replace…
00:25:14 Manjeet Rege
Yeah, So those kind of jobs are at risk of displacement.
But if you look at the entire number of jobs, there have been a lot of studies, number of jobs that will be created by AI and number of jobs that will be lost and there are a number of studies that say that by 2032-2035 overall you will have more jobs created compared to the number of jobs that are lost.
When you look at certain white collar jobs could be a radiologist, a marketer, a business analyst. HR analyst, even now I believe the number is something like 70 to 80% of the resumes are looked at an AI, are looked by an AI model first.
00:26:01 Manjeet Rege
So an HR person did not have that in-depth AI knowledge, but that person is getting exposed to AI and that's what I meant by you're getting exposed. So, you need to then become knowledgeable about using tools that underneath might be using AI, but on the surface you know you don't need to have an in-depth knowledge about AI.
And then people, certain kind of jobs that might get displaced, then one has to kind of re-skill them and make sure that you know workforce is re-skilled and able to get back doing jobs that are able to combine their years and years of domain knowledge along with this new added AI skill.
00:26:47 Mae Macfarlane
And what do you recommend to kind of do that? Like what would it be that companies would have to invest in or just us as a people?
00:26:55 Manjeet Rege
So when we talk about companies redefining themselves as an AI first company, it is not just about changing the technology overnight. It's about changing the DNA of your enterprise.
The biggest asset for companies is their workforce, the people who worked there for years and who have lots of knowledge. So investing in them and re-skilling them, which a number of companies do, that is the easiest way for helping people transition into newer roles and by them transitioning into newer roles, people are also able to contribute more with this new added on skill and that ends up benefiting the company itself.
00:27:44 Mae Macfarlane
Yeah, that's something that I always kind of have thought about, especially with the you mentioned the HR thing.
You know, AI does take a look at resumes and there's buzzwords they kind of always look for. But when you're an applicant who, as I kind of mentioned earlier, is coming from a literature or a biology or business background and you're applying for AI jobs to get into data science and all this stuff and you don't have any of those things on your resume.
What is kind of being done about those people that are falling through the cracks in like the job search because of AI's interference? Is there anything to support that information or?
00:28:17 Manjeet Rege
It has been shown that some of the AI HR tools have a bias towards male applicants, and since that was disclosed, there is a lot of work being done where that is getting minimized.
So that is one way through which -- because the problem is at the tool level, and if it is developing a bias, it has to kind of overcome that AI bias and elimination of that is an active area of research that is being done.
As far as an applicant is concerned having your information out in multiple outlets and showing how you differentiate yourself from others is an important part. And that's what I tell my students as well. In our curriculum, we have a capstone which is an optional capstone. Students can take any other regular course.
And in data science what happens is hiring managers want to look at what you can do or what you have done independently, as opposed to your grade.
And that is, I think, becoming true for a lot of other disciplines as well, where if you can develop a portfolio of all of your projects, put it on, let's say GitHub, which is a repository, where an employer can look at that and say, “OK, you have done all of this,” so the conversation becomes more about the work that you have done as opposed to your transcript, which is similar to 100 other people. I think from an applicant standpoint as well, becoming aware of putting your information out there in a lot of other resources, having a strong LinkedIn profile, GitHub profile, if you are in the data field is absolutely essential.
00:30:12 Mae Macfarlane
What is one question, or one topic, that when you're talking about it, get really excited about, but that the average person or people in general don't really share the same level of interest?
00:30:23 Manjeet Rege
So, I'm really excited about how AI can generate new content. So, one aspect of a human, like creativity. You know, we can write, we can draw, we can paint. And in the past, quite a lot of years, in the past 8-9 years now AI has been able to make advancement. It generates content that looks so really…like it is able to write stories for us. It is able to get inspired from a particular kind of painting and apply that style to another painting. It can listen to a particular artist, get inspired from that, and generate new music.
00:31:06 Manjeet Rege
So I'm really excited about the applications of that. Many times people do not realize how deep AI is and the wide applications of that.
We only get to see in our real-world the different applications around us, like the automatic thermostat or a self-driving car. Just like humans AI adjusts itself along the way.
For example, when the pandemic hit, a number of the AI models were taken by surprise because the credit card transactions patterns changed. People were not shopping online as much, right? So, before pre-pandemic, if you made a lot of purchases, the AI model would say you know this is very unlike what you had done in the past. But now with the pandemic, a lot of transactions moved online.
So typically, those would have been flagged as a suspicious transaction, but the I had to go back and that this, this is where the human comes in, right? So, the human would come in and say that we need to retrain our AI model so that this is being considered as normal and not a suspicious transaction.
00:32:21 Merrie Davidson
I do have a question about human learning and machine learning and what the differences are. And when you talked about the car and the child, the computer gets tons of examples.
00:32:35 Merrie Davidson
And then the child probably gets fewer. But my assumption is that the child is taking into different considerations as well, like it's self-moving, it's blah blah blah, it's blah blah blah.
So for the generalized information they have as a human, they're putting more into the car and.The AI is just like looking at pictures and seeing only the shapes.
I guess then that in order for the computer to learn more like children or act more like children, they need to have that generalized information that they can gather from different places.
00:33:11 Manjeet Rege
So when you talk about the differences between human learning and machine learning, if you think about learning in the first sense, you're capturing features and you capture the features, you analyze those features, and then you make a prediction.
00:33:31 Manjeet Rege
At a very basic level, the feature capturing about quality of a human versus a computer is different when you provide information to a machine learning model. 15-20 years ago, the features that it captured were things like the texture of the image, color of the image, but not the semantics. So not the meaning.
00:33:57 Manjeet Rege
So, for example, sample red flower and the red car -- at a color level, they are similar. But we know they are semantically different. So, now with some newer features it is able to also capture a lot of other things, but still it lacks far behind a human level understanding.
So that is why the generalized intelligence and machine learning level of AI is far off from each other, so also as humans we are able to transfer some of our prior learning, some of that context into a particular task. Where an AI can struggle with that.
00:34:41 Mae Macfarlane
I guess kind of going off of that, what would we have to do to create that next level of understanding and comparison?
00:34:48 Manjeet Rege
So I think that's a very difficult question -- whether we'll ever get there or not. Because as I said, we've had some advancement in certain niche or narrow paths.
00:35:02 Manjeet Rege
But you know in in the past when you look at how the AI as a field has evolved from the 1950s, there have been effort to do something like that.
If you...So,there was an effort called building export systems. So the idea behind expert system was “
“Let's interview a doctor and let's take all of that information. Basically, code it up, put it into a computer. Let's take a K12 feature. Ask a lot of questions, put it into a computer, and then let's kind of then build a hierarchy or let's combine all of these together so that when you pose a question, we'll send it to the right export system and have that question answered.”
But that didn't go too far.
00:35:52 Manjeet Rege
It had a lot of promise on paper, but implementing something like that is a lot of challenge that you extract all of this information from, let's say, a school teacher.
But if you have to extract that information from a doctor, you have to again start from the beginning and you do all of that same process.
You're not able to capitalize on your first expert system, whereas a human learning is so quick and it is able to adapt so quickly to the changes, right?
I mean if you again going back to the pandemic, we were able, all of us were able to adapt to the changes so rapidly on a number of levels. That is working from home, teaching from home, learning from home.
That is something that an ML or an AI falls behind a lot. Although I am optimistic, I'm also realistic in terms of where AI will take us. There are a number of capabilities, limitations of AI.
When you look at somebody, you can quickly figure out the person's intent. You have a police officer on the street, let's say saying “stop.” “Stop your car.”
The hand is raised.
You might have a hitchhiker, a different intent.
You might have your friend. And again, raising your hand and saying “hi” to you.
So those three gestures are similar at a feature level, at an image level. But when you look at that, you could quickly figure out the intent.
But if you give that to a image classification AI, it will not figure out the intent.
So unless we get to that stage we are not close at all to getting to that generalized intelligence.
00:37:36 Mae Macfarlane
Thank you so much for joining us today,Manjeet it's great to talk to you.
00:37:40 Manjeet Rege
Thank you so much. It was fun.
00:37:41 Mae Macfarlane
To learn more about Doctor Rege's work, you can find them on the Saint Paul campus of the University of Saint Thomas.
Professor Podcast is brought to you by Saint Thomas libraries and made possible with funding from the College of Arts and Sciences.
I'm your host, Mae McFarland, a 2022 graduate. The producers and library staff are Merrie Davidson, Andrea Koeppe, and Trent Brager. Thank you so much to our guests and to you, our listeners.
0 Comments.