Curiosity AI in Education Conference, Interview with John Sokol
Technology and Education: A Conversation with John Sokol
A Curiosity Conference Session
Hosted by Gordon
Introduction
Gordon: Welcome to today's curiosity conference session. My name is Gordon and today with me I have John Sokol. Our topic of the day will surround technology and its relation to education. So first, John, would you like to tell us a little more about what you do and who you are?
John: Me, I'm sort of a self-taught computer expert. I've been running around in Silicon Valley for the last 30 years involved in countless startups. My focus lately has been on pure robotics and robotic education and things related to that.
Current Work at Roboterra
Gordon: So right now you're working at Roboterra. What's one project there that you're proud of or you're excited to have worked on or working on?
John: Yeah, so there's a couple projects we worked on over the last year and a half, two years. One was a telescope, and another one is a snake that uses reinforcement learning to move. Also we got involved with a cat and a few other robots like this, so I hope to finish that up as well.
Teaching AI Principles Through Robotics
Gordon: Well I think that these robots seem to be able to teach kids or to educate people in a way that is unique and to teach them about the fundamentals of artificial intelligence and all these programming skills. What do you think about that?
John: Yeah, there's several simple AI examples out there that I've managed to put together that can teach AI principles in basically like two pages of C code that a student can understand. Simple things like learning the rules and how not to fall off a table on a little driving robot, for example, all the way up to reinforcement learning where we've got an 8-servo articulated snake that will learn how to move on its own and figure out what the best movement patterns are using currently using rules from Mathematica.
Gordon: Are these implementations something like a simplified version towards a young adult or teenager population, or is it the same sort of coding or structure that would be used in an actual implementation on a professional level?
John: Well, AI is an enormously large field. So the goal here was to have a robot that you could build over the course of a class or two and then in the course of a one-hour period, apply the learning algorithm and watch it go from flailing around the table to being able to effectively move itself. And be able to make it so that we can apply that to almost any kind of robot and have it figure out how to best move itself to the situation without us programming it.
Visual Intuition in Learning
Gordon: That is cool. I guess that kind of relates to how you can combine AI in such a way that makes it visually intuitive for the students to understand the inner intricacies. How do you think that sort of intuition plays out? Would that help them in the future or in what aspects would that help them in further learning of STEM and future knowledge?
John: Again, the goal was to be able to teach a couple of aspects of AI and machine learning from the perspective that a lot of AI and machine learning is really an extension of data science, which is statistics and data analysis, and then be able to apply this to robots. In the case of reinforcement learning or other learning algorithms that we can teach them, this in the course of an hour, watch it, and get some intuition for what it's doing and basically demystify what's going on under the covers—something that's small enough and simple enough.
AI in Education Delivery
Gordon: So as I remember, there's one part of using AI in the product itself. What about as Roboterra or in the field—are there any ways of using AI to help learners learn better? As in using it in the teaching program or solutions itself?
John: You know, Roboterra is not doing this directly, but with a lot of online education, I think that machine learning and AI can be used to optimize the curriculum—basically learning what type of course material is most effective for different types of students and be able to identify what students, let's say, are maybe more audio learners than visual learners or hands-on, you know, different learning types, or ways of connecting them with the content better.
Potential Biases in AI-Driven Learning
Gordon: Would you think one implication with that sort of identifying strengths and weaknesses—would there be a potential that AI would create like a self-reinforcing or self-fulfilling prophecy? For example, if it identifies a student as a visual learner, would it be a possibility that even though the student might not actually be a visual learner, because of how the algorithm might guide the student, they become better at visual learning and create sort of a bias or deviation in their strengths and weaknesses?
John: I think we're a long ways off from having this kind of confirmation bias or some sort of self-reinforcing thing where it sends students down certain paths. Certainly we're afraid of that happening with our police and other institutions where suddenly it's like, "Oh, you've had problems in the past so we're just perpetually red-flagged." Our credit system seems to do this to people—you've got bad credit so we're going to give you harsher scores for every little minor infraction, whereas somebody with good credit, you're late on a payment, it's nothing.
So yeah, these systems are currently being used in banking and credit for years—these kinds of machine learning algorithms constantly looking at odds and sizing people up for insurance rates, for example.
Gordon: Right, so yeah, another problem in the educational field—it's not a problem yet, but we call them filter bubbles in Google or Netflix recommendation engines. So yeah, I kind of see it as there's 20 different ways we can teach you trigonometry. Which set of videos or instructional materials is probably going to be most effective for the student? I don't think there's going to be a lot of room to really deviate too far out of the lines though, right?
John: Right, especially with more fundamental knowledge, it's hard to actually do so.
AI as Teaching Assistants
Gordon: But yeah, speaking of the ways that you can teach students, I'd assume a part of these algorithms might also be focused on giving feedback for problem solutions. In the case that perhaps if there's like 10 ways to solve a problem—which I guess is more of a common phenomenon in programming—how would an algorithm go about providing feedback for the varieties of combinations of ways that a student could approach a programming problem?
John: [AI is] not very good at finding what's wrong so much as making predictions or guesses. The prime example is reading handwriting on a letter—is that a three or five with sloppy handwriting? In the case of applying this to student learning, certainly regular programming algorithms—if-then statements, you know, they get the right answer for one plus one, it's pretty straightforward. But figuring out if the student's having trouble, how to address it, or gauging what the best approaches are...
My feeling is this is more of—think of it this way: with AI you start building bots. Now these bots aren't necessarily physical things, these are software entities, but they assist you. You're trying to achieve a certain goal and suddenly you're able to create a software entity that will look for certain stock behaviors or alert you for certain things, kind of like Amazon Echo.
So these are very useful for AI, but actually getting to the point where the student has some understanding of what these tools can and can't do and how to apply them—I see them as more of amplifying a teacher's ability. Currently one teacher is split up against 30 students, but with AI it's like having a room full of really dumb teacher assistants helping. So there could be effectively some interactive one-on-one with the student going on in a way that the teacher wouldn't have time to correct them and go, "Well no, this is what you did wrong."
The Changing Role of Teachers
Gordon: Yeah, instead of facilitating the more low-level sort of learning process, while a teacher may help with more higher level or more complex or hard-to-define issues with learning.
John: I would imagine the teacher ends up doing more of a supervisor role rather than hands-on. At some level, yeah, the first pass is like go watch the video, use the online tutorial or some interactive aid for that particular coursework, and then the teacher is monitoring the process.
Gordon: Yeah, because I mean knowledge deliverance doesn't really need a teacher to be present. I mean they could just read textbooks.
John: There's a scene in one of these Star Treks with all the Vulcan students in their little computer pods learning. Yeah, I kind of see that—every student's going to be sitting in front of a laptop and most of the coursework is going to be there.
Social and Emotional Learning
Gordon: Yeah, in that case would there be a need for social emotional learning or recess time and whatnot? That sort of devoid of relatively lesser human interaction within the classroom might have other effects.
John: Well, I think our understanding of what classrooms are is going to change pretty drastically too, because the computer industry has really changed the way that people have managed to work together and learn and share information and knowledge in a much more efficient fashion. I see that culture now propagating out to other fields and industries. The biotech people are starting to pick up on it and take advantage of GitHub and other tools, and many other fields as well.
In the end, basically people in the future are going to be learning for the rest of their lives. It's not like you go to school, you get your K through 12 and you're done, maybe you get some college with some module beyond that. The reality is things are perpetually moving, and every time you sit down for a new project there's a lot of learning that's going to have to happen continuously because things are constantly evolving and changing.
Getting the students in the mode where they can sit down and you can point them in any direction and they can start learning on their own—"Oh okay, here's how gravity and acceleration work to do a cannonball game or something to this effect"—and how can we provide them tools and AI assistance where suddenly they can have the equivalent of a hands-on mentor that will happily explain to them all the physics and gravity and things, rather than them having to fish through 100 hours of YouTube videos and a thousand pages of physics texts.
AI as a Personal Learning Assistant
Gordon: Sort of set up where AI has a role, because I mean I know we have analogies like everything is on Google now, and with everything on Google it's sort of like an extension of our brain that helps us sift through information faster than we could process.
John: Yeah, but Google doesn't digest information before giving it to you. I think of it this way: imagine you can hire a personal assistant. You go, "Okay, we want to figure out how to draw a circle. What's the math function for that?" And instead of having to go through and fish through all these pages of websites and theory, somebody who understands this stuff can come back to you and go, "Okay yeah, there's three different algorithms for doing this, and depending upon what you're trying to achieve you're going to use one of these"—Pythagorean theorem versus sine versus cosine—and why would you choose that, and something that can sit down and explain this to the student so that they can be more functional hands-on.
But at the same time, there's not a lot of point in rote memorization because in the future those AI are going to be there for you actively helping you. So really it's becoming more of learning how to take advantage of these tools to empower you. With today's tools, you can do stuff that a guy with a PhD couldn't do 30 years ago. Now a 12-year-old kid can pick it up and slap together some face recognition on a Raspberry Pi. The code's there, you just learn how to download the examples and boom, you're rolling.
All of the knowledge to learn is there. Now the question is, some kids figure out how to pick up those tools on their own and learn how to read the docs and the manuals and get up to speed, and a lot of kids don't. So getting them to the point where they have that self-motivation, self-learning ability, I think is going to be a key to surviving in the future, because the AI really should be a tool to speed up your ability to learn.
Educational System Reform
Gordon: Right. Well yeah, I mean that seems like—especially with us in the States—the educational system would be very hard to prepare us for that kind of forecast, I guess.
John: Well, you know, yeah, there was some example where they're basically training factory workers for 70 years ago still. Where it's like, "Okay, we're gonna have to write or develop something to do X," and here's a whole bunch of things you don't know about, and you're gonna have to read up and learn about X, Y, and Z to accomplish your mission. And that's all like learning new subjects without a teacher sitting there holding your hand.
Gordon: Yeah, I mean with regards to that, what was your experience like going through the state's educational system a few decades ago, and how do you feel? What kind of changes would you make to the past system or the current system if you could do so?
John: Oh my god, I did not have a good experience at all in my public education system. Yeah, and so I was more focused on making money and learning and doing stuff with my computer and ignoring the academics of the school system. So later, having to come back up to speed on basic English skills, for example—I look back at my writing from 30 years ago and yeah, I really look illiterate. But my education was very lopsided because I was very unhappy with the academic school system at the time.
Gordon: Right, yeah, but I guess with the self-initiative you could pick the English back up whenever you felt like there was a need of it.
John: Well, I have in the course of 30 years of doing business—business plans with misspelling and bad English—at some point it's just doing it over and over and over again that has forced me to clean up things. But even just communicating on social media where a lot of it's all typed, for example, has kind of forced me to improve my spelling and English.
Gordon: All right, so as soon as you put some stuff out there, you've got several friends on you going, "Oh, you misspelled that."
John: Yeah, we're there. That's a better way to learn instead of a teacher giving you feedback—your friends, which I think gives more motivation for things.
Gordon: Certainly peer pressure provides a lot of motivation.
John: Yeah, having a group of friends... Well, I'm seeing my son right now trying to do these programming hackathons, and the thing he's finding the most difficult is they don't know how to collaborate. So I grew up and I was involved in the publishing of the first open source operating system, the 386 BSD, that eventually the Linux guys came over and they had their thing. They borrowed a lot from what we had posted, but the most interesting thing is they were more successful in organizing the group and the community...
The Importance of Collaboration in Education
Gordon: Yeah, and distribution and stuff where there was a lot of chaos when we first posted, for example. Right, um, and learning how to communicate and interact as a group and sort of on board with other remote teams and work remotely—like especially in this day and age is a big deal. You know, it's a skill that very few people have the ability to do that where they could just work at home. You know, even though there's literally nothing at a desk other than the computer at their office, for some reason, they just don't have the skill to let's say communicate effectively over email and other means to make up for not physically sitting next to someone.
John: Yeah, and even then when you put people next to them, you know, just learning how to cooperate and work as a team is difficult. So yeah, if the schools taught that ability for students just to collaboratively work together, um, that would be an enormous benefit for the next generation of information workers. You know, I don't care what their marketing, stocks, or anything right, let alone programming and AI, uh, where you end up with larger and larger teams collaborating. Absolutely, yeah.
Gordon: I guess that's kind of hard to teach at school, I guess especially about the remote working part. You can't just have students learning at home and trying to create an environment. Yeah, I guess that's with everything well.
Different Approaches to Education
Gordon: So there's different schools I've seen. There's like the Waldorf schools, for example, where they've got a very open creative environment, very different than the public school system where you know, and now I'm looking at what the kids were being pushed through and they've got like packets which were completely mindless. You know, the teachers spend an enormous amount of time organizing and grading these packets of stuff back and forth. You know, right, um, and a lot of that should be this should be a program you know on the computer doing most of the heavy lifting. You know, even like grading uh essays and and writing and stuff like this could be done by AI easily.
John: Yeah, or like, yeah, I could get some objective scores where humans could swear like the creativity part or whatnot. Yeah, so you know, they had the AI that was uh looking for plagiarism. Yeah, and the higher universities so when you submit a paper they put it into the engine and see if you plagiarize anything. Um, would that be AI or is it like comparing text to text?
John: Well, I guess again, it's kind of quasi-AI because again, they're using these statistical means of going well, you know, your essays, you know, a 92 percent match to this other document or something. Right, and so they're able to catch more than just okay, we just rearranged a few things and then you know we swiped a paragraph here, you know, change something they'll catch that. Right, so it's more than just like a direct you know pattern match. Oh yeah, I'd imagine because you can change synonyms and maybe that will catch it too. Yeah, yeah, that's true. Yeah, I don't wonder like if essays could be scored objectively, it's kind of like asking AI to critique art. Know it's extremely good. They've got it now where they they can you know throw a bunch of papers in in the grades and then and then throw a paper in and it gives you what it thinks a human would have graded it.
John: Right, you know, and the beautiful part is they've got like adversarial learning and other techniques now where they took the one AI and pitched it against the other. So it's like, oh, okay, so now let's come up with completely garbage papers that pass with an egg you know on the grading machine and and you know it completely fakes it because you know the grading machine tends to score well if you've got certain let's say word combinations or something right. You know, it's just like pastes this document so there's a famous uh one of the AI challenges is you know, there's like a recognition algorithm for let's say a dog and now they give it like some completely random picture that has nothing to do with the dog. You know, there's like one white pixel in the right place and then suddenly sees a dog. You know, the AI. Yeah, you know, they had to use an AI to find that that hole right to to defeat the other AI. But the fact is that you know, they were able to to fool the algorithm in such a fashion that was so overtly gross to like a human would look at and go, what the you know, why did that AI fail like you know, yeah.
The Future of AI and Education
John: Um, yeah, I guess that's I guess that's still a progress we have to work with for future. Yeah, well, so in robotics we call them corner cases where you've got it working for most things and once in a while you know, yeah, things will go wrong. So you know, I have a little cartoon from like in the 70s or 80s it said, "To errors, humans are really foul things up takes a computer." You know, and it's true. You'd send a job to the printer and all sudden you know 500 pages of random gibberish would come spewing out automatically. You know, that's like you know, well, if you're not there watching right, you know, the humans would make a mistake but the machine would just kind of keep going and going and going and going. Right, you know, entire pack of paper before you noticed something was wrong. Right, oh yeah, actually coming back to the essay example, I was just I just thought of like the case of Amazon and how the AI was biased in uh shifting to resumes because I mean, as in the same because like it kind of it replicated human biases without being biased itself. So so this is where a lot of the AI comes in is is it's black magic, you know, kind of voodoo as opposed to science in regards to how you present the data to to train it and so it's a lot like like making grass and you say you know, it's so easy to lie with graphs and statistics. Um, and the reality is is when you're training a machine a lot of it's you know, you're you're altering and shaping the data and presenting it to the machine to learn and if you're not careful in your methods, you don't end up with the successful outcome or you end up with all these these biases that you didn't expect at the time.
John: Yeah, yeah, but that's the more problem that data set than actually the algorithm itself because it only replicates the choices of humans and humans well, but again, the data set was cultivated and groomed by a human. Yeah, you know, the famous one is they were trying to separate tanks from cars and and and all the tanks and pictures and cloudy days and all the cars and pictures and bright sunny days and the machine learning algorithm was able to separate the tanks from the car images because anything with a bright sunny day had to be a car and and then when they tried applying it to other things suddenly they you know, they realized it didn't have any connection to recognizing the vehicle at all, it was looking up right. Yeah, yeah, I guess that that's a problem with their scientists and all that.
The Evolution of Machine Learning
Gordon: So you know, yeah, and things like this but you know, we're learning to overcome them and you know, I'm surprised that they're teaching it in schools professionally now. Well, I mean, just machine learning, you know, I mean, 30 years ago it was a really obscure thing when I started. Right, then how do you start with machine learning 30 years ago? What brought you in?
John: Oh, so yeah, I would hang out with the smartest kids I could find back in the day and and so we're always you know trying to like one-up each other. Um, and so you know, neural networks started happening and we had you know a couple hundred lines of C code with a simple neural net. I was just doing you know a couple layers with multiplication and things and producing you know experimenting and playing with these things. You know, the stuff we're doing today with GPUs is essentially just you know massive versions of that simple algorithm from 30 years ago which is you know we didn't have more than a couple of megabytes of RAM back then and now we've got you know gigabytes with unbelievable horsepower. You know, I'm just just starting to notice my browser tabs got more power than the half million dollar SGI workstation I was experimenting with at Stanford 30 years ago. I mean, in terms of like GPU and computation and everything right. How much it's probably got more more compute power than the Cray YMP that we were renting time on to do the computations before we sent it to the SGI even. Like, you know, in JavaScript it's it's just amazing, crazy.
The Future of AI and Computing
Gordon: Yeah, so how we can also just connect to cloud computing like on browser types as well just with more computing power to both disposal. Yeah, yeah, so anyway, yeah, but so I mean if like 40 years ago neural networks has already started like it's an infant state within these 40 years are there any things that are like do you see being like infants that will grow to become the next big thing or like the really the next sort of like technological sort of big wave of advancements in like Silicon Valley.
John: Well, I mean, there's a bunch of big waves that have been happening um, I mean, well, you know, so right now we're just at the point where we can process live video images and actually do useful things with them that we couldn't do before. Um, and that's opening up a whole universe of opportunities. All right, are you back? All right, that's bad. Right, yeah, so we were talking about live image processing and well, what does that come out or what is that right like what does that signify in terms of so um for example, uh, at this point we can take a live video from a video camera and and work out the camera's movement. Um, right, we can recognize objects in the room so so you know the kind of trivial examples and there's even some ones that run browser attached to that open my webcam, I don't recognize the cat versus a chair or other objects in the room. It's not great but you know, it's okay. There's an algorithm called segmentation that will separate like you know people from from other things and try and label them um and and these algorithms are getting faster and and they've got a really nice framework now where they can compare them you know how much compute power versus accuracy and and quality for example um and and you know so as computers get faster and faster this is going to be just you know normal this kind of uh ability for for even a low end machine um to have an awareness of the environment it's in where before computers really were oblivious to what was going on in the room around it. Right, uh, you know, so it would have a sense of context if you will you know of the situation and what's going on. Right, um, yeah, you know, I think that's going to change things dramatically. Yeah, I can imagine yeah, I was I guess I was researching on buying like a Roomba just a few months ago and like I was I was pretty amazed at how like the progress has gotten since like five years ago where there seems to be some level of intelligence and it's like routine or obstacle aversion but I guess like if life inspires like if these processing really gets better then maybe the algorithms and its capability could be much better than right now or even a new type of cleaning sort of robots could appear rather than just before itself so so you keep mentioning they're very limited yeah.
The Power of Algorithm Improvements
John: So you know, I'm on the subject of algorithms and the improvement there's something called the Commodore 64 demo scene and so basically the Commodore 64 is you know a fixed computer they made 30 years ago it's like a historical piece at this point um the amazing thing is the demos that people have been able to write now that run on them you know because of the advancements and algorithms uh compression performance uh so just you know same hardware as 30 years ago nothing's improved at all but what they can do on the Commodore 64 now in terms of music and video and animation is just astonishing you know that and 30 years ago just just we could never dream that we could do that with the machine um so you know that's just an algorithm improvement um and algorithm improvements tend to be the larger shift you know so there's Moore's law where the computers are just getting faster because the chemistry is better and the more transistors they could just throw more hardware at it right but then you know if you realize oh I didn't need to do all these extra multiplications because we can do it some other scheme and reduce the math my god the performance enhancement on that is is tremendous.
John: Wow, so we're seeing a lot of that with the AI as well is not just the hardware improvements but but algorithm improvements and understanding things so it's not uncommon for them to throw enormous money at enormous hardware to try and and get let's say a technology going but once that you know people were in there they go oh okay we can reduce and simplify and and and this works just as well with you know one tenth the horsepower and and before you know it you know um these algorithms are getting more effective with less compute power. Wow that that I didn't know about that. I mean 30 years ago I can't imagine running like modern services on such I mean a machine that predates me. Yeah, so so yeah I mean if you think about it now you know what we're gonna have mobile in our pocket is is you know effectively what what like a half million dollars with a hardware by issue today which is you know incomprehensible right it's like a server room. Yeah, yeah, so so the ability of generating photorealistic images and animations with all kinds of intelligence and understanding the the sensors that are attached which we can't even imagine but I would imagine it would at least have a full depth understanding of of all the structure and depth and what's going on around you spatially. Right, yeah, I mean yeah and it would be able to do that for almost for free for free like oh yeah the cost you know the cameras are going to be you know a dollar a piece that the compute power you know is going to be five bucks worth of CPU on the phone right they'll just naturally be able to do a 3D reconstruction of like every room you've ever stepped into with it right all right that that's practically for free yeah. I mean I mean basically yeah you know I mean right now what's pushing the limit on a phone like this today you know eventually people are going to figure out and post open source you know just like some kid will just like download it and and okay yeah I did this as part of my you know kindergarten show-and-tell oh that'll be scary yeah yeah as there's a feature it's right ahead everything well I mean I'm watching the bar you know for these things come down so even you know what we were doing at the robotics club 30 years ago people were able to do with an Arduino in a pie now you know things considerably more advanced you know and I still see people struggling over how to put two two wheels on a box with a battery still you know the wheels the motors the battery you know haven't changed all that much in 30 years but with the compute power and not only that the documentation on the internet where you can find you know hundreds and hundreds of examples of web pages and blogs and videos you know that will teach you step by step on how to do this um the problem is is is it's an overwhelming amount and and figuring out which one do you want to give a student versus sending them down this this overwhelming okay there's a hundred ways to go with which one and suddenly that you know they'll end up being paralyzed by too much information yeah then maybe then you have to rely on a recommendation algorithm to give you the one it thinks you need yeah yeah well I think we're up to time with our session today and I mean thank you John for being here because I mean amazing and I believe for me to learn about all these advancements in technology.
No comments:
Post a Comment