Josh Tenenbaum receives MacArthur Fellowship
Tenenbaum emphasizes collaboration, exploration of various aspects of computation and cognition
Brain and cognitive sciences professor Josh Tenenbaum PhD ’99 was awarded a MacArthur Fellowship Sept. 25. Tenenbaum leads research initiatives in the Computer Science and Artificial Intelligence Laboratory, the Center for Brains, Minds, and Machines (CBMM), and the MIT Quest for Intelligence. The Tech spoke to Tenenbaum about his work and the field of computational cognitive science.
This interview has been edited for length and clarity.
The Tech: What does winning the MacArthur Fellowship mean to you?
Tenenbaum: It’s obviously an incredible honor. Informally, it’s called the “genius” award, but most of the work I do is very collaborative, so to me it means recognition for being able to do great work in collaboration with amazing people.
TT: How will you use the grant to further your research?
Tenenbaum: One of the things I might use the grant for is to seed new, far-out, crazy projects. The other thing is to support programs that increase access to and participation in our field for people from underrepresented groups. We have some very good programs in our department that don’t get nearly as much funding as they deserve, and I feel lucky that I can use some of my funding to support these programs.
TT: Computational cognitive science seems to be a fairly new field. What are some of the most exciting research discoveries you’ve made at MIT?
Tenenbaum: “Discovery” is a funny word because in a lot of areas of science, it isn’t so much discovery, but a kind of understanding. We study common sense. While a machine might, in learning a concept, require hundreds of thousands of examples provided by a human engineer or user who has some dataset, a child learning a concept might learn from only one or two examples.
Imagine growing up in a city where there aren’t any horses typically walking down the streets. The first time you see a horse, it’s a really exciting event, and your parents might say, “Look! A horse!” And that one instance might be enough to learn what that word means, to be able to figure out which things in the world are horses and which things aren’t. You can get that concept, learn a word for it, and be able to generalize from just that one example. Those are some of the amazing things that people learn, and we’re able to some extent describe how they do that using computation and mathematics.
TT: What work did you envision yourself doing when you first came to MIT for graduate school? How has that vision evolved over time?
Tenenbaum: When I came to MIT for grad school, I imagined myself doing exactly what I’m doing now. My father was an early artificial intelligence researcher, and my mother worked in education and studied how kids learn. I was lucky to have some great mentors as an undergraduate. I did my PhD in the same department where I work now.
A couple of things have evolved. There are some specific projects that I was really interested in 25 years ago, which were too early to work on then. Before graduate school, I worked on a project in “intuitive physics,” trying to understand how the human mind is able to think about and imagine complex physical situations.
Then, starting about ten years ago, some great folks in my lab started working again on intuitive physics using a technology called physics engines — tools that people use to make interactive physics-based video games. So, many techniques which had been developed over the intervening years, both software and hardware, allowed us to build these models.
All the interest that has arisen in artificial intelligence in the last few years has also taken on increasing importance in what we do.
TT: What are the greatest challenges you’ve faced in bridging the gap between computer science and human cognition?
Tenenbaum: The first gap that comes to mind has to do with cultural gaps, and gaps of language. The communities that study these topics are often different groups of people who are used to thinking and talking in different ways.
The fields of computer science and cognitive science were born together and grew up together. A lot of what we had to do in bridging these fields is develop ways of talking to each other — to figure out ways to use the math or the formal systems of computation to think about how the mind works, and to convince psychologists or neuroscientists that these tools are actually the right ones.
TT: What inspires you in your work?
Tenenbaum: The early mentors I had. I was lucky to work with one of the great cognitive psychologists of the twentieth century, Roger Shepard, who was a professor at Stanford, when I was an undergrad. [Shepard and] my PhD advisor here at MIT, Whitman Richards, were inspiring figures for me, and I think about the lessons that I learned from them on a daily basis.
In the last few years, I’ve been especially inspired by scientists who study children’s development. My specialty is computation, but scientists like Elizabeth Spelke or Susan Carey, who both used to be MIT faculty but are now faculty at Harvard, are some of the great developmental psychologists of all time. Some of my closest colleagues at MIT, Rebecca Saxe and Laura Schultz, also study the brains and minds of young babies and children, and their work has deeply inspired me. I’m incredibly lucky that I get to work with all of them collaboratively on these projects.
TT: With the founding of the Schwarzman College of Computing, there has recently been a great emphasis on the ethics of technology. As a leader in the MIT Quest for Intelligence, how do you approach the ethics of AI?
Tenenbaum: Many researchers at MIT, including myself, are still figuring out how to do that because we’re not trained in ethics.
The kind of work that we do is potentially related to machines that can affect the labor market. David Autor, Daron Acemoglu, and many others at MIT have studied for a long time the impacts of automation and recent advances in AI on the general picture of jobs. And that’s something we’ve been trying to talk to those colleagues and others about — to understand how the work that we do can make people work more efficiently, more creatively, more satisfyingly, and not less.
Another clear ethical issue has to do with who’s paying for the research and what priorities we’re spending [the money] on. MIT is doing some much-overdue reflection about those issues. That’s something the faculty has to take very seriously.
TT: What are your future plans for your research? What are some potential applications of your work that you look forward to?
Tenenbaum: The thing that I’m most excited about right now is a project that we’ve been working on as part of the Quest for Intelligence and CBMM. We’re working on the idea that we might be able to build machine intelligence that grows into intelligence the way a person does.
What I’m most excited about going forward is building that bridge between the fields of cognition and AI around how children learn to think. I’m especially interested, going into a somewhat farther future, in the computational grounds of what words mean and how people put them together into questions and answers. We have on the AI side … machines that can generate natural looking English text in amazing ways, but they’ll babble and go off in completely crazy or boring directions that no human who has ability with language would ever do. We are interested in taking the kind of understanding that we’ve been achieving in basic, prelinguistic kinds of common sense and seeing how those might carry over to understanding how language works.
TT: Do you have any advice for members of the MIT community who would like to get involved in AI or cognition?
Tenenbaum: Take the opportunity of being at MIT to learn very broadly and deeply about all the different aspects of cognition and AI. There’s so much attention on the field right now that many students are only looking at a very small part of the picture — in particular, there’s a lot of interest these days in deep learning methods and papers that have been published since 2015.
One of the great things about MIT is that there are lots of people who have been working on cognition and AI for decades using a range of different techniques.
So, take the opportunity to learn from different kinds of people of different ages, especially from the older faculty or people who take approaches from different areas of science. Be aware that wherever you start, it’s just one small corner of the space. Be open, explore, and really try to listen and learn from a broad range of people and perspectives.