Among Terry Sejnowski’s many interests has been a focus on education, specifically how we can use what we know about learning in the brain to improve education. A little more than a decade ago, he helped found The Temporal Dynamics of Learning Center to understand how the elements of time and timing are critical for learning, and to apply this understanding to improve educational practice. In the second part of my Eye on AI interview with Terry, we talked about that work, when we can expect findings to be productionized for students in the market, and how artificial intelligence is shaping the future of education generally.

Add speed and simplicity to your Machine Learning workflow today

Get startedContact Sales

TERRY SEJNOWSKI:

The Temporal Dynamics of Learning Center was one of six science of learning centers funded by the National Science Foundation ten years ago. It was a $30 million, ten-year project involving 12 institutions, 50 investigators and 50 fellows, working together collaboratively on a wide range of projects. Our center focused on machine learning and neuroscience, putting those two together. And I'll just give you a couple of examples of what we did.
There was a robot that Javier Movellan put together called Rubi that he brought into a preschool classroom of 18-month-old toddlers. The purpose of the project was to try to see if we could get toddlers to interact with the robot. The robot wasn't very sophisticated, it just sat there, but it had very expressive eyes that moved around and it had hands that could pick things up and had a ‘teletubby’ touchscreen with things that the kids could press. And it was able to play music.
The idea was to interact with little kids, so here's what happened. The kids are running around, you know, they have very short attention spans and so they run over to the robot, ‘what is this?’ And the boys grabbed the arm and pulled it off because they're testing things. And so, he went back to the shop to repair it. Rather than put in an industrial strength arm, which he could have, but that would be very dangerous, what he did was put in a pressure sensor. And so, when that arm was yanked the next time, Rubi would cry; the boys backed off and the girls hugged it.
So, this is social engineering. What is it that humans respond to? How do you get humans to behave? How to interact with them to get their attention and hold their attention? If you have a teaching robot, that's what you've got to do. You've got to be very interactive. And so, we discovered many things. Like common attention - this is very well known among people who study child development. When a mother and a child are together, they have a common attention. When the mother points to an object, the baby will look at it. Most species don't do that. They don't have this common attention. Rubi was programmed to have common attention. So, if a little kid pointed to the clock, Ruby would look at the clock. The kids love this. They could do this for hours because they had some control over this creature.
And then we started incorporating learning, new words and language learning and so forth. This was a great experiment. We learned so much from it.
We were afraid that the teacher was going to feel threatened because you hear about AI taking your jobs. But the teacher loved it. Why? Because for the teacher it helped her with crowd control. One of the biggest problems when you're teaching toddlers is just keeping everybody in line. And Rubi was a way of doing that, helping her as an assistant. It was assisting her to be a better teacher. And that's been the experience we've had all along. Whenever we do something that looks like it's going to threaten jobs, it basically helps people do their jobs better.
Now the second project I undertook with Barbara Oakley, who is an engineer at Oakland University in Michigan. We put together a MOOC, a Massive Open Online Course, which became widely popular, called ‘Learning How to Learn.’ Our goal was to try to help students become better learners, taking advantage of what we know about how the brain learns.
Barbara would give a lecture about how to solve a problem, like exam anxiety, a mental block, procrastination. And then I would give a lecture about what's going on in your brain and why this is happening and why her advice is going to work. I'm giving brain lessons, and at the same time, Barbara is giving them practical lessons and we use a green screen so that we could have things in the background, pictures of neurons, things flying around, just like the weatherman.
We also used humor to try to get people's emotions involved rather than just having a dry talking head. In any case, it became the most popular MOOC. Now over 3 million people have taken it. I get fan mail every day, people in 200 countries, ages 10 to 90 and it really is having an impact. And this is I think where we are headed. We can use the assistants like Alexa, for example. You can imagine Alexa being programmed so that it interacts and helps educate individuals.
It's been well established that by far the best education is when you have one-on-one instruction with an adult who is a really good educator and a child who the adult understands, and can help get through mental blocks and so forth. But that's very expensive, very labor intensive.
So, if you look at what deep learning has been successful at, it's been the machine-human interface: speech, vision, language. Those are the channels that humans communicate with each other through. And that's where we've made the most progress. That can be put into use now, immediately, in order to bring this to the student. Instead of sitting in front of a computer and pressing buttons, now the student can talk to Alexa or the tutor or the system or whatever. Now that can be a dialogue. It's going to be much more efficient. It'll be much more natural and it'll make it more fun to have someone that you can talk to, rather than a static computer.
Amazon actually has a competition for people to write programs for Alexa. There are programs out there now for Alexa that are being built.
Education, I think, is going to be the killer app for deep learning and the reason is, it's so important. It's such a problem. We have terrible problems in the U.S. where our educational systems have failed us K through 12. This has long-term consequences. It's going to go on for generations. So, we have a new technology now that can be delivered. My MOOC is an example of the first wave of that. This is one-way, but eventually it'll be two-way. It'll be one-on-one, and that's going to make education so much richer and so much better for everybody.
It's not going to get rid of teachers because you still need humans in the loop overseeing things and making decisions about when does a student go onto the next block and so forth. Very complex decisions that have to be made.
Our Temporal Dynamics of Learning Center grant finished in 2018. What we are doing now is going international. We've organized meetings with other countries and we’ve inspired science of learning centers in Australia and Brazil. Many countries now are following the lead that we took 10 years ago.
And now we're getting some of the big foundations interested, like Chan-Zuckerberg and Gates Foundation. These are multi-million dollar powerhouses and we're reaching that point now where, with the right international cooperation and funding from the major foundations who are in this sphere of trying to help education, I think this could be a real worldwide effort to create a much better learning environment that is based on the brain, on how learning really works and also delivers teaching through the most powerful channel we have now for communicating, the Internet. All the pieces are falling into place.
Through this center we thought that we were going to be bringing science to the classroom. But, the classroom brought data to us. We learned an enormous amount.
Delivering what we learned, however, was incredibly difficult because of all the barriers. There are gatekeepers at every step. There are 12,000 school districts in the U.S. and if you needed to take some software into a classroom, you're going to knock on 12,000 doors. And then if you get past those doors, you have a problem with unions. Do teachers want to learn a completely new curriculum that you'd need to actually deliver this stuff?
We're academics, we don't know how to do this. You'd have to start a business and it would take many decades because our educational system is so frozen. It’s a huge, multi-trillion-dollar industry. So, Coursera taught us something. It taught us that the way that you get into people's homes is to bypass all the gatekeepers and go directly through the Internet. Anybody with an internet connection, any place in the world now has access. The lesson is that you don't try to reform the system. That's not going to happen. But using the tools now that are available with machine learning and the internet, you can jump over that.
Parents, of course, want the best for their kids. And parents are figuring out that they can get much better lessons for their kids through the Internet than they're getting at school.
That's going to accelerate. You have a personal teaching assistant that's coming out of 10 million of these smart speakers that people have in their homes. You’ve just got to know how to use it, to program it. You've got to get deep learning in the loop. So that's where I think things are headed.
How long will that take? Well, in a sense the hardest part has already been done, which is building up the infrastructure. The infrastructure is there and now it's going to be companies going in with software and building up better and better interfaces. It's not easy. We're dealing with very complex problems that kids have when they're learning. So, you have to really understand something about being a really good educator at the same time that you're being a good engineer. But that will happen. It's already happening.

This post is adapted from the second half of Episode 31 of the podcast Eye on AI. Find the full recorded episode here, or listen to it below.