By Ashok Goel & Ida Camacho
The Association for the Advancement of Artificial Intelligence (AAAI) and Squirrel AI Learning announced the establishment of a new $1M annual award for societal benefits of AI. The award will be sponsored by Squirrel AI Learning as part of its mission to promote the use of artificial intelligence with lasting positive effects for society. The new Squirrel AI Award for Artificial Intelligence for the Benefit of Humanity was announced jointly by Derek Haoyang Li, Founder and Chairman of Squirrel AI Learning, and Yolanda Gil, President of AAAI, at the 2019 conference for AI for adaptive Education (AIAED) in Beijing.
Established in 2014, Squirrel AI Learning Intelligent Adaptive Education by Yixue Group is the first artificial intelligence company in China to apply AI-powered adaptive learning technology to K12 education. Squirrel AI Learning products use a model that combines artificial intelligence and human coaches to provide students with access to individualized and affordable high-quality education. Although the focus of Squirrel AI Learning is on education, Li insisted that the award will recognize AI innovations across all disciplines. The establishment of this award aims to inspire the AI community and draw attention to AI that can benefit humanity. This new international award will recognize significant contributions in the field of artificial intelligence with profound societal impact that have generated otherwise unattainable value for humanity. The award nomination and selection process will be designed by a committee led by AAAI that will include representatives from international organizations with relevant expertise that will be designated by Squirrel AI Learning.
In this interview, Richard Tong, Squirrel AI Learning Chief Architect & GM of US Operations, discusses adaptive learning, challenges facing AI in education, and the Squirrel AI Award for Artificial Intelligence to Benefit Humanity. This interview has been edited for clarity and brevity.
Interview with Squirrel AI Learning’s Richard Tong
Hello Richard. To begin, can you tell us a little bit about Squirrel AI Learning?
Derek Li is the founder at Squirrel AI Learning. As a serial entrepreneur, Derek started his first company after graduating from college and all three companies that he started are education related. While he was the head of his second company, Derek discovered that he had about 28,000 teachers but had an annual turnover of about 8000 teachers. He realized that as teachers gained experience, they were more likely to leave for a variety of reasons. For instance, teachers may decide to open their own tutoring companies or jump to smaller institutions where they can be more in control.
It’s a huge problem for these after school programs because the quality of the teachers is very important. In after-school programs the parents have control, rather than the school district, or the teacher union. This makes the brand of the institution less important than the brands of the individual teachers. Parents and students place greater importance on individual teachers’ quality and reputation.
Derek began with the question, “How do I maintain the quality and also retain institutional knowledge to be reused?” Again, looking at different ways, he made a lot of tries. He built a large product team, curricula, and tools that can be reused. He also worked with content providers as well as other research and development companies to standardize the best practices and training.
Around that time, he came to the US to look at some of the US companies. One of the companies he visited was the company that I used to work for, called Knewton. From his visit to Knewton, he was impressed that there was a company that could actually use AI to get a lot of teaching knowledge, use data to validate the knowledge, and reuse such knowledge in a general way.
He also analyzed another company, called Kumon, which had a tremendous impact on his approach. Kumon is an after-school learning company. Their model is what we call manual adaptive rather than AI adaptive. In this model, students attend school, they get a worksheet. Based on each student’s performance, he or she might get a different worksheet. The teachers don’t actually teach, they just give students practices based on their progress. Kumon’s model is very useful, especially when you’re teaching very fundamental skills that require a lot of practice and really targeted.
He looked at these two types of models and noticed that the Kumon model is very old school, but it does not rely on the teachers that much. He also examined Knewton, where instruction recommendations and learning paths are recommended by an AI machine. And so, he formed a new company combining them and he went out to start Squirrel AI. Originally, he started with some of the people from his earlier company. These are the more avant-garde folks who wanted to do something different and something more impactful.
The Squirrel AI model is B2C. We have franchisees, but we control everything. We’re combining three key components for an AI solution, the AI engine itself, the content and pedagogy, and the operation and service. We want to use AI to teach. The model is not about the teacher anymore. The model has shifted the center of the learning universe to students. It’s all about one-on-one instruction, feedback, practice and personalization.
That’s the Bloom’s Sigma effect.
We shift the focus from the teacher as a deliverer of the information to the student who is learning. If you were to compare the traditional education model and the one-on-one teaching model, it would be like comparing a bus to a taxi or an Uber.
In the traditional model, the biggest time waste is when a teacher is teaching at the pace that the student is not efficient, and this is almost the case all the time. Even you’re teaching to the middle. Each middle person is different. For example, at AAAI conference, we go to these big lectures, but the most efficient interactions are one–on–ones. When I go into the lectures, I already know some of the content, yet I cannot follow some of the other content, and I get lost easily. At these big lectures, I would say that 10% of my time is efficient where I’m actively learning, thinking, and getting things out of it. Even with a good professor, I may be listening to things that I already know or things that are too far ahead for me. Even when I cannot catch up, she continues and moves on to the next topic. That happens all the time in the traditional classrooms.
This is where AI comes in. With AI, we can afford one-on-one teaching for everyone. In the past, the barrier is that the teacher is a limited resource. It’s a resource that we cannot afford to give to every kid, especially those from poorer background.
Currently, we have an AI and coach hybrid model. We still have a human coach that is providing the emotional support and filling in the gaps. The majority of the time, teaching can be done very routinely, repetitively and consistently by AI. However, the student could get stuck or have trouble understanding the AI-provided instructions. That’s where the human coaching comes in.
It’s a complete spectrum for the age group that we’re talking about in all subject matters. But we are not a test prep company. Traditionally a lot of adaptive companies come from test prep. For test prep, the learning goals are very clear, and it is very easy to do optimization. You can say “If I want to optimize test scores, then these are the things you have to know”.
And this is an ill-defined problem.
For us, we are one step further. Our learning goals are just synchronizing with what they’re learning in school, so it’s much easier. It is a little bit more difficult than the test prep but is still a lot easier to optimize than teaching creativity or computational thinking where the goals are so difficult to define and you cannot build objective functions very effectively.
So how would school measure success?
The Squirrel AI model follows a process where you start from diagnosis. Next is recommendation, feedback, and collaboration. In the future, we hope we can also include motivation.
In this process, the two key steps are diagnosis and recommendation. We continually diagnose students as they’re learning. We track keystrokes so that we know where you are, how far, how fast you’re solving a problem, and other relevant knowledge. Then we calculate the probability that you’re going to be able to solve the next question, and use that to diagnose the student learner profile, so that we build this continuous learning profile. We feed the profile to a learning map and our objective functions.
You look at the knowledge graph and say, “In order to learn these concepts, you probably need to learn that some other concepts first. Also, these two ideas are similar. Maybe you can try this.” We call it a GPS for learning.
Going back to the analogy of a taxi versus a bus. Taxis must be driven by a human driver with a great deal of mapping information. With GPS, you don’t need to be a super-knowledgeable human driver anymore. Instead, you can have a decent driver that does not have a lot the mapping information, as long as the GPS is available.
The notion that we can build the student profile is a very open, very hard problem. You mentioned leveraging content maps, learning trajectory, and learning progression to able to optimize the learning for every individual student. In a way that has a dream of AI in education for the last 40 years, right?
It is. However the fundamental question is still difficult because we still don’t know how humans learn, we’re still struggling with that.
A lot of time when we talk about kids younger than seven years old, mostly using system–one learning. After nine years old, they are using a different learning algorithm altogether. The principles that we are looking at—Bloom’s two sigma, forgetting curves, zone of proximal development— are fundamental. We could go the whole nine yards to solve all the detailed nuances. But also, even to apply the basic stuff is already going to have impact. The big thing is that you’re competing against a very inefficient alternative, the traditional system. Even if you’re just giving students the right difficulty level, you are already so far ahead.
So I really like that because instead of solving a big problem, which none of us fully understands, what you’re saying is we are here and we can make a delta difference is already a win.
Yes, it is already a big win. We have demonstrated that in efficacy studies in what we call Human Versus Machine Competition. These Human Versus Machine Competitions were conducted both for publicity and evangelism, but also because we wanted to see that we were able to beat human teachers and the traditional model every time, hands down.
Do you think there is a point at which a human teacher becomes less necessary?
Again, let’s go back to the taxi and Uber analogy. If you look at taxi drivers when GPS software and the Uber system came out, the total number of drivers exploded. There are so many more drivers because there are so many that are enabled by the smart tools.
What we see is that, as a society, we benefited from more personalized services. In the past, we didn’t have so many people taking taxis all the time, but now everybody’s using Uber or Lift.
At Squirrel AI, we are seeing that same trajectory. Right at the beginning, you’re going to see a lot of teachers who were not previously qualified enough actually can make a difference. In our case, a substantial number of our coaches are just out of college. Traditionally these people cannot really be very effective, even as one-on-one tutors. Now they can be because we’re expanding the capability for them. That is helping students to get much higher benefits at a more affordable price.
Over the longer term, we’re also looking at another shift. In the past, if a parent is good at everything and knowledgeable in every field, the teacher will probably have easier time teaching. A lot of the good students have very good parents at home to provide one-on-one help. This parental support can be hugely different depending on family background. But with AI, we can almost say that every parent can also become so much more effective.
So, this is a very important point. If we can make education more accessible and affordable, then it helps make it more inclusive and equitable. And everyone has an access to it.
Yeah, one of the narratives we had was about gender. If you look at the past, we see that even good teachers have stereotypes. My daughter often tells me that her teacher doesn’t think that she’s good at math. You know that the teacher does not mean to be biased, it’s just human nature. But when we have an artificial AI tutor, they don’t exhibit gender or racial biases.
They look at what you know and what you don’t know. They will tell you that you got it wrong. That is much more objective. It actually reduces the gender gap because when you look at a lot of the gender gap, it’s not because the girls do not perform well in STEM, but it’s their surroundings. It’s the people who they think are authoritative figures, or who they look up to that matter. We are leveling the playing field and reducing bias by bringing high quality AI-powered support.
Okay, so how would you know that you succeeded in five years, five years passed by …
There are several thresholds we want to cross. One is that from a business model point of view, we want it to be become more scalable and also more sustainable. This means that we really hope that AI can become an important factor and ingredient for success, and not only just for us, but for the whole industry, not only in China but also across the globe. We hope that personalized learning platforms using AI could become mainstream and fruitful. And I think it can be regarded as a success when AI education becomes 50% of that world education.
But in 10 years, we really feel that AI should be part of the whole education system. We are dealing with the after–school programs, but we want public school education to also get something out of this whole revolution. We’re talking about the humanity benefit because if you think about in countries such as India or China, the majority don’t have high-quality education at all.
We can make a difference. Since we can save so much in terms of human resources, teachers can do much more meaningful things and bring much bigger impacts on kids’ lives. That’s what we regard as success.
How did Squirrel AI think of the Squirrel AI award?
This comes back to the question about how we think AI should be applied. As we’re challenging these very difficult problems, we have found that AI needs four things: good business models, a lot of data, algorithms, and computing infrastructure.
The biggest part is actually about the AI talent, who is developing the algorithms, who is deciding what type of algorithm should be applied, who is using data efficiently, and so forth. When we look at the field, AI is already widely used. But when we take a closer look at AI in education and nobody really cares. If you look at the literature over the last 10 years, there has been not much progress at all. We are still stuck with ACT-R and SOAR. Because of the buzz with neural networks and deep learning, everybody is focused on image recognition and auto-driving.
For example, look at the problem Squirrel AI is addressing in education. If you make a little progress, it’s huge for society and huge for kids, especially the kids in poorer neighborhoods where they just don’t even have good teachers at all, let alone personalized teaching. If we can apply AI to this field, it is a huge win for humanity.
While Squirrel AI focuses on education, you have chosen to make the Squirrel AI Award open to innovations across all domains. Tell us about this decision and how it relates to your motivations for sponsoring this award.
We want to shift some attention and apply AI to the benefit of humanity. Our chief AI officer, Tom Mitchell said, “I want to do something that would have a lasting impact on humanity.” One of the reasons that we wanted to set up this program was that we wanted to attract good researchers, especially top AI researchers, to pay some attention to things in AI that may not be as fancy as image recognition or translation. Tom originally suggested focusing on fields such as healthcare, education, environment protection, and smart cities.
Have you considered how this award will shape the average citizen’s view of AI?
With Squirrel AI learning, parents do have apprehension to a lot of the non-human parts. This is where award can help us do a better job of educating the whole public. Right now, we hear a lot about the bad aspects of AI. Biases and deep fake have become the narrative and we don’t often hear about the good. The award is not centered around profit or smarter weapons or smarter surveillance tools. Rather, we want to shift the narrative to AI successes for humanity.
Why did you pick AAAI as a partner?
The first reason is that I would say it is probably more coincidence than anything. This came up as an idea that originally when Yolanda Gil, AAAI President, went to one of our conferences. We were initially thinking of organizations that have a big footprint in terms of the research community and already have success in terms of promoting these contributions. That’s a good place to start these types of conversations.
We also look at the way that we want to make the biggest dent. We don’t necessarily want to create another Turing Award. I think what the field is lacking is that there is no good AI award such as the Turing Award and the Nobel Prize. Also, there is nothing that is really focused on the humanity aspect. Yolanda and Tom helped brainstorm how we can combine these two things and who we should partner with.
Another mandate is that they wanted it to be very international. So that it’s not just US–focused or China–focused or European–focused. We want to bring good organizations together. So, we created a three-organization core with AAAI as the leading organization, along with EurAI and CAAI. Again, we are a China-based company and would like some home support. The goal is to create it for an international community. When we look at the humanity problem, it is not a single country issue. It’s really about the whole human race.
For example, healthcare is something that is indiscriminate. Education is the same thing, it’s indiscriminate. If you improve education, anyone can benefit. That’s where we are focusing on. Yolanda has been really, really supportive. She really helped to drive the planning of the award and nomination process.
Briefly what can you tell us about the award and nomination process?
In the Award Committee, there are seven people– Yoshua Bengio, Tara Chklovski, Edward Feigenbaum, Yolanda Gil, Xue Lan, Robin Murphy, and Barry O’Sullivan.
We definitely want this award to be a AAAI award first and foremost. Even though we are very honored to have our name associated with it, we really want this to be all about the bigger goals than what a single company is all about.
How will you know that you have succeeded with the Squirrel AI Award 5 years from now?
If we attract maybe 1% of the PhD students getting into AI to think about their impact on humanity in a beneficial way. That’s huge. I feel that when we look at the four fields that we mentioned—healthcare, education, environment, and smart cities. In any of these, if we can see that AI researchers are thinking not just about profit or competing in big AI competitions, but about humanity, it would make a dent. We are still just a small part of it, but it takes a whole community to pay attention.