Training the Next Generation in AI #153 | Embracing Digital Transformation | Intel Business
Hello, this is Darren Pulsipher, chief solution, architect of public sector at Intel. And welcome to Embracing Digital Transformation, where we investigate effective change, leveraging people process and technology. On today's episode, Training the Next Generation in AI with special guest Pete Schmitz.
Pete, welcome to the show. Thank you very much, Darren. Glad to be here.
Pete and I have a sordid history of working at Intel together. But Pete, tell us a little bit about what you've been up to and your background. Sure. Glad to. So for many years, I worked at Intel as an account executive, helping out the Department of Defense, figuring out which of the Intel technologies could be useful to help them advance their mission. So after a very wonderful tenure working with many people, I retired from Intel at the end of last year, and I've been focusing on continuing the volunteer work I've done as part of the Defense Industry Consortium called FCA.
That's the Armed Forces Communications and Electronics Association. My role in the San Diego chapter has been to help STEM students write science, technology, engineering, math in the kit mathematicians to develop a love and pursue a career in STEM that might wind up aiding U.S. based defense, but it might not.
So it's been good work, and I've taken the interest that I've developed over the years in artificial intelligence and begun to teach the students because in their work as robotics competitors, especially at the high school level, they need to use autonomy in their competition. So I thought perfect place to educate them on this technology for their competition and for use in the future. So I think it's wonderful.
I mean, this volunteer, because you're building you're building the next generation of entrepreneurs and scientists that are going to extend this incredible field of artificial intelligence that we're starting to see now. I think I think that's great. Thank you for for doing that. Yeah, that's terrific. They're just amazing kids. And, you know, they have such incredible interest and capabilities now. That's been wonderful to see them absorb.
I mean, I've worked with many amazing technical customers over the years, and I can just feel these students being on the cusp of being that next generation of implementers there. Oh, that's great. That must keep you young. Exactly. Exactly. That's awesome. All right. So you've you've had some time to focus on artificial intelligence over this last seven, eight months that you've been retired. What have you found? Well, I've found that there's a real need for people to understand a little bit better how this technology works.
That was what was driving me to understand so I could explain it. Right. There's a lot of FUD, a lot of hype concern. Is it going to take over a future? Is it going to be embedded in our brain? So having been exposed to Intel's efforts in artificial intelligence and having a head start, I thought, well, I'm going to dive in and learn myself how it works and what better way to figure out what I know, what I don't know. Then by explaining it to the students? Well, that's, you know, that's what they say. The best way to learn something is to teach it.
Exactly. So. All right, so what? So what have you found? Well, what was kind of eye opening for you? Well, let me just give you a representation of who these students are. And that might help to, you know, shed a little light on on, you know, the appetite. So can you take this screen? No, not yet.
Did you hit share? I think I did. Let me try it again. All right, now it's coming through. Okay. All right. There we go. All right.
So this is a group of high school students in this case from San Diego High School. But this robot that they're standing in front of operates against a set of rules in a competition. And part of this is to use autonomy and the autonomy portion in this case uses computer vision. So I thought, well, I'm going to help these students with an with a possible architecture in their autonomy portion. Right. They use QR codes for positional information to determine where is the robot to help navigate it, to pick up an object, etc..
So I thought, well, I'll explain to these students how this capability is currently being used in the defense world. So I took for them and described an example of the DARPA D Hunter unmanned surface vehicle. This is the Defense Advanced Research Projects Agency's continuous trail submarine. It's a trimaran.
You know, it's a vehicle to trail submarines continuously. Is the idea. It's a 132 foot vessel that has an outrigger on each side that's completely unmanned and it uses radar and cameras to implement itself piloting capability. It can go up to 10,000 nautical miles.
Wow. And operate up to 21 knots. So when a camera is looking out from the bow of this device, of this vessel, it's very similar to the students robotics and how they spot threat obstacles, things that they need to investigate or or steer away from is based on computer vision, based on artificial intelligence. So I took them through the idea that in artificial intelligence, you are gathering a lot of data right? This could be structured data, it could be unstructured data, but it's all being used to make predictions using a model based on training data.
And this capability, the concept of machine learning has been around for a long time, but it's evolving so quickly that they have been born just at the right time in history to be able to ride this exponential increase in technology as it's coming together because there's a convergence of the latest capabilities of math, data, silicon and programing. They're really unleashing all kinds of world changing possibilities. So are you seeing that these high school students, are they stepping up to it? Are they able to understand? Because this is really complex stuff, right? It is it still too complex for a high school student to grasp it? Or the neat thing about these teams, Darren, is that they want to be very inclusive in gathering team members of all capabilities. So of course you're going to have the programmers that can grasp and implement this technology, but you have mechanical engineers to design the system. You have marketing students to let folks know about the capabilities of the team. You have operational folks to help implement the logistics.
So I'm finding that the computer programing students on the team are already well into this capability and very open and passionate about how they can use this technology, especially the math portion. I thought that was particularly interesting to figure out, Gosh, I cook a lot of linear algebra and calculus in college for my engineering degree. How is that used? It turns out that in artificial intelligence, linear algebra and calculus are widely used because you need to use matrix multiplication.
For instance, in the case of convolutional neural networks for computer vision to figure out and detect what is in the image that I'm detecting or seeing in my camera. So finally, we can tell our kids you're learning this because you're actually going to use it. How many times have you heard, I'm never going to use this. All the time. Why? I think it's awesome because these kids actually get to learn a couple things.
First off, that what they're learning in school is valuable and they're learning, I mean, artificial intelligence. They're standing on the shoulders of 50, 60 years of research in this space that they they can just take for granted. Oh, yeah, it works fine. I, I think it's wonderful for them to have that without having to go through all the pain and turmoil that the industry has already gone through.
That's really true. And to open their eyes to the fact that, like you just said, it's been around for 50 years or so that folks have been trying to learn from data. But now with the silicon capabilities and the programing capabilities that exist, it's being accelerated to the point where structured data that had to be labeled very laboriously. You know, that's a great way up to this point. But the doors are opening up to this new generation of artificial intelligence where unstructured data can be used on, you know, some AI supervised learning can be used to to remove the the burden of I'm going to label one image at a time with is that a raccoon or a monkey to determine in my picture what is this item that I'm seeing? So so because advancements in technology and silicon, frankly, I mean, the silicon is a lot faster now.
So we can do a lot more. Right? These students are able to take advantage of the of these really high tech things much easier. The barrier to adoption is a lot lower than than it has been even four years ago.
That's really true. I mean, take the case of the students. They have a certain budget that they cannot spend more than the allocated amount to build their system.
Right? Otherwise, it would be like the world of competitive sailing. You just throw money at it to build the best you can. It's just not fair. So the students have a budget where maybe they've got a couple of hundred dollars to spend on the accelerators that they might use.
It really forces them to learn the intricacies of A.I. to figure out what really is the latency that I need to navigate autonomously and get an accelerator proportionally priced right, or maybe use just the capabilities of the CPU to do that. So it's a well, a good real word exercise. Yeah, I like that because it's a real world, right? We don't have unlimited budgets, even though we'd like to have unlimited budgets, we don't have them. And there are power constraints and size, right? You can't hook your autonomous robot up to the cloud and have all your data processing happen in the cloud, right? That that just doesn't work as well.
So I it sounds like a great program. It's fantastic and it really is a platform and an opportunity to open their eyes to, okay, you've got this project you're working on, but as you pick your path in education and decide what do I want to do when I graduate from college, what I want to study to help get there, the types of questions that you can answer using AI and deep learning are much broader, right? In other words, in the autonomy world, you might just be asking, Yeah, what? What is that object out there? Is it what I'm looking for? What should I do? But in the broader scale, you can go from just detection to maybe you want to decide. I want to categorize a bunch of images.
I might want to predict an outcome. Maybe I want to make a recommendation based on the data that I'm seeing. Maybe I want to generate new contact content.
And these can be used in a variety of applications in health care, for instance, you can use it to maybe detect cancer, assisting an oncologist with identifying an object that maybe would be missed or should be looked at more deeply or in the retail world, maybe you want to, you know, target ads based on what a user has in their shopping cards. Correlate with what you have in inventory correlated with the highest margin product you want to present to them in their shopping exercise. It sounds like it's on air is kind of unleashing a whole a whole myriad of new business cases, business use opportunities in instead of it getting rid of jobs. It actually may create several, several new ways of doing business.
That's really true. And it's not just for the folks that are the developers, the scientists or the engineers, because to answer these complex questions might take the insider the perspective of a business person or a marketer or a data analyst. So the doors are opening for all kinds of work roles to use A.I., not just the technical folks that can digest the nuts and bolts. So so this brings up an interesting question around generative AI because you re you retired right? When Jenny I just kind of blew up November of 2023. Now we're in 2023, now 2022, November 2022.
Jen I opened Open Eyes. Chad came out and took the world by storm. What you what have you seen over the last six, seven months? What has that done with the kids that you're teaching and in the industry as a whole? What's your perspective on that? Well, it's been fascinating because first I offer them a comparison between, well, what was being used before that convolutional neural networks were mimicking the functionality of the brain. Right? They had multiple layers. They had perceptron all these layers and perceptions were connected.
Input was fed into this convolutional neural network. Wait. Their assigned feedback was generated until you got the output that you had hoped for. You compare it to the input and calculate the loss.
Right? That's the neural network using labeled input. Now the and that was great for image classification, object detection, maybe speech recognition and that uses the, you know, the mathematical functionality of a matrix multiplication. A lot of pre-trained models exist out there, so folks are not starting from scratch. They can use transfer learning to take image recognition capability and decide, okay, I'm going to keep a couple of layers.
And now change subsequent layers to detect that particular item that I'm looking for. That's different than what it was trained on, right? But with Transformers and and Generation generative A.I., the real breakthrough was in finding patterns between elements mathematically that exist over a long range. You've got a long stream of data coming in, but you want to draw information from this whole stream of data, this whole image. They use positional encoding.
So the model that came out, the mathematical model uses normalization layers, residual connections, and it's resulted in a number of amazing new capabilities. Right. Not only the generative Pre-Trained Transformer model, the GPT, but also Bert, the bi directional encoder representation from Transformers, their vision transformers, the vanity of the model and the one of the latest ones is called the Megatron Turing Natural Language Generation. Now the real interesting thing is that these models would not have been possible to train and use without the advent of the latest capabilities. Right? It takes millions of dollars, billions of petaflops the chain to train these models that can now be used. And both are wanting to use these capabilities for things like drug discovery.
If we can predict the shape of a protein and its consequent functionality, we can accelerate the time to produce and personalized new medicines, for instance. It's just fantastic. So it so it is a game changer.
That's what I'm hearing. Yeah. To this do do you do. You think that I know there's always a watershed moment in different technologies even if general even if GPT is not the most complex or most capable technology out there, it changed the world because no ability. Absolutely. Yeah. Yeah. People are now thinking beyond, oh, AI's doing object detection because that's what people were thinking. I was was good for and, and more people are interested which means more money is being dumped into it.
How, where do you see things put on your crystal. You know get your crystal ball out and talking to the students. The art of the possible.
So where do you see things going in the next five years as far as A.I. and the way it can change education and or or any industry? Well, I asked the students to take a look at I give them their first dose of business insight. This may be really cool. It might make a wonderful video game, but when you get out into the real world, you have to ask yourself, does this offer the capability to make a company money or save a company money? And that, of course, as you can see, the promise and the hype of this has driven up valuations in the stock market. Folks expect that the future earnings of companies are going to increase because AI is going to improve their ability to reduce costs. Will it replace employees? I don't think so.
I think it will make them more productive. It will free them to be more creative. Will it allow companies to make more money? I think you're going to see a combination of companies that can either use the technology to improve their revenues and their income, or it will allow customers or companies to be born that will disrupt those companies. They will be born without the shackles of existing processes and infrastructure and can use AI to build new business models. So examples of those business models could be in encoding, right? I've got students that write a lot of code. I asked them how would it be if you could use AI to generate the first pass of your code and then you could complete the code.
Very interesting to them. I mentioned life sciences use in drug discovery or in chemistry. Translate Passion.
Boy, How about the language gaps that exist in cultures around the world? Real time translation is possible to use possible to do with these large language models. Well, in fact, I don't know if you know this. I have another podcast that does the weekly news in Digital transformation, and I produce it in six languages. Oh boy. And I don't do translation.
I have a generative, I do the translation and I have native speakers in those countries, review it and then read for me. And before when I looked at doing it, it was going to cost me substantial amount of money to do the translations. Now literally, it takes an hour's worth of someone's time.
That's a native speaker to read through the script and read it. And it's it's one it's, it's opened up new doors. So I see exactly where you're going with that right.
So the other thing that I let my students know is where are the centers of excellence that exist in developing these new innovative technologies? Of course, they're with companies like Intel that develop silicon, that have accelerating instructions built into the CPU's or that have GPUs available for data center or end user. But they also exist in some of the software companies. So for instance, the the trans former model itself, I let them know that came out of a Google Brain initiative in 2017.
The concept of stable diffusion, which is generating high quality, detailed images that came out of Stanford. Maybe you want to go to Stanford and do some work there and then other technologies that are on the cusp now are something called the neural radiance Field and Nerve. This is going to be really applicable in autonomous navigation and robotics.
You'll be able to generate 3D content from 2D images coming out of industry. So all kinds of places that I've tried to raise their awareness or this work can be done. And further using. AI, I think that's wonderful because you're showing them AI's not over, it's just starting. It really is. And over the horizon there's this concept that's being talked about called artificial general intelligence, very different than just artificial intelligence.
But the general intelligence portion is the idea that down the road you might be a part of developing or using something called a generalizable embodied agent. And this is an agent that can adapt to new conditions. It can handle high dimension or complement complicated data distributions using underlying technologies like neural operators to understand multiple domains, like perhaps you want to mix image and text and scientific all together.
I think that's going to be on the next wave of innovation. Very, very disruptive. It absolutely.
In fact, it feels I'm sure you remember the early nineties when the Internet started just taking off. I mean, the Internet had been around for how many years, decades before the nineties when people started getting in their homes and e-mail was starting to take off. It feels to me like this is in the same boat as that. It really is. And and you're going to have some friction, some some barriers to adoption. So when you come up with a great new use case, you've you've gotten your data, you've identified the problem you want to solve, you figured out which model am I going to start with? That has got to mix with the infrastructure, the capabilities, the people, the culture that already exist in the company that's going to use it, right? You and I have worked on great projects that involve enterprise, compute architecture and the architecture that's going to host this A.I.
Innovation is much is similar in that it's going to have a goal of how am I going to develop this, test it, store it, iterate it, you know, check the value, and then in the end, you know, how accurate is. I've got to be able to explain it when this model is running. I've got to match that to my environment. If it's on a very low power autonomous system, I can't just put a big 300 watt GPU out there. I might need a low power FPGA.
So all these things need to be considered when you're going through the development cycle of getting ready to deploy that model. I think that's where the ecosystem is very interesting. Now you've got all kinds of folks out there that are figuring out how to specialize from an innovation perspective. Am I going to be a company that provides tools across the whole spectrum, like the cloud service providers, right? At least have to match with the infrastructure. So maybe you're going to piecemeal your data analytics portion, maybe you're going to piecemeal your experiment management portion to try and build and evolve this environment where these students, once they've been educated and trained, can go to work and be really productive in delivering that capability. Well, well, that's another question I have, is we need to we need to, in the industry, make sure that we have the environment so the new workers that are coming in can hop on to these new technologies, can be innovative that we don't.
I love I love the word when you said earlier shackle them down with old processes and old ways of thinking. We need to unleash the power of of of these new engineers, business leaders, marketing people to to think outside of the traditional boxes that we that we've been dealing with for for so long. That really is true. And I think that's why you've seen the rise to prominence of some of these companies, like like a snowflake, for instance. This is going to be a specialized capability and product to be able to manage your data, maybe a tacked on or hops work, for instance, enables you to do feature management in your development pipeline.
Your operations, as you and I have worked together before on DATABRICKS, but has a whole lot of other company there, like heavy AI or data robot. Maybe you want to have a tool to label your data. You might use something like label box or scale or appen. All kinds of technologies exist out there that have to be put together to fit with your environment, your capabilities, and the objective that you're trying to accomplish.
So it sounds to me like we're not being replaced at all. There's a lot of work to be done still. I think that's, you know, the onus is on us, Darren, to figure out ways to enable our our students, our customers to really solve these hard problems, whether it's in robotics or video analytics, health care, public sector, public safety, media, entertainment, genome X, autonomous vehicles or Internet services all offer terrific opportunities to solve amazing problems. It's really an exciting time going forward. I think so, too. Very, very excited.
I'm not scared at all. I'm excited. But I also know this You either evolve or die, and so I'm trying to educate everyone, jump onto this bandwagon because this is going fast.
So exciting stuff. Pete, thanks for coming on the show. I love the perspective of the next generation of of leaders that you're training up.
Thank you for obviously helping helping the industry as a whole. Was that my pleasure, Darren. Thank you for having me on the show. Thank you for listening to Embracing Digital Transformation today.
If you enjoyed our podcast, give it five stars on your favorite podcasting site or YouTube channel, you can find out more information about embracing digital transformation and embracingdigital.org Until next time, go out and do something wonderful.
2023-08-27 19:29
