The December 2022 issue of IEEE Spectrum is here!

Close bar

Lab Revisits the Task of Putting Common Sense in AI

New nonprofit Basis hopes to model human reasoning to inform science and public policy

5 min read
ai hand and human hand touching pointer fingers

The field of artificial intelligence has embraced deep learning—in which algorithms find patterns in big data sets—after moving on from earlier systems that more explicitly modeled human reasoning. But deep learning has its flaws: AI models often show a lack of common sense, for example. A new nonprofit, Basis, hopes to build software tools that advance the earlier method of modeling human reasoning, and then apply that method toward pressing problems in scientific discovery and public policy.

To date, Basis has received a government grant and a donation of a few million dollars. Advisors include Rui Costa, a neuroscientist who heads the Allen Institute in Seattle, and Anthony Philippakis, the chief data officer of the Broad Institute in Cambridge, Mass. In July, over tacos at the International Conference on Machine Intelligence, I spoke with Zenna Tavares, a Basis cofounder, and Sam Witty, a Basis research scientist, about human intelligence, problems with academia, and trash collection. The following transcript has been edited for brevity and clarity.

How did Basis get started?

Zenna Tavares: I graduated from MIT in early 2020, just before the pandemic. My research had been around probabilistic inference and causal reasoning. I made pretty complicated simulation models. For example, if you’re driving a car, and you crash, would you have crashed had you been driving slower? I built some tools for automating that kind of reasoning. But it’s hard work to do in a conventional academic environment. It requires more than one graduate student working on it at a time. So how can we build an organization focused on this somewhat non-mainstream approach to AI research? Also, being a little bit burned out by my Ph.D., I was thinking it would be great if we could apply this to the real world.

What makes your approach non-mainstream?

Tavares: The mainstream right now in AI research is deep machine learning, where you get a lot of data and train a big model to try to learn patterns. Whether it be GPT-3, or DALL-E, a lot of these models are based on trying to emulate human performance by matching human data. Our approach is different in that we’re trying to understand some basic principles of reasoning. Humans build mental models of the world, and we use those models to make inferences about how the world works. And by inferences, I mean predictions into the future, or counterfactuals—how would the world have been had things been different? We work a lot with representations, like simulation-based models, that allow you to express very complicated things. Can we build really sophisticated models, both for commonsense reasoning but also for science?

Sam Witty: The application areas that we’re particularly interested in, and I think have been underserved by a lot of the existing machine-learning literature, rely on a lot of human knowledge. And often, scientists have a lot of knowledge that they could bring to bear on a problem. One main technical theme of our work is going to be about hybridizing, getting the best of classical approaches to AI based on reasoning, and modern machine-learning techniques, where scientists and policymakers can communicate partial knowledge about the world and then fill in the gaps with machine learning.

Why aren’t causal methods used more often?

Tavares: On the one hand, it’s just a really hard technical problem. And then two, a lot of advances in deep learning come because large companies have invested in that particular technology. You can now just download a software package and build a neural network.

Witty: I think a part of it is the kinds of problems we’re trying to solve. Think of the application areas that large tech companies have focused on. They benefit from vast amounts of data and don’t rely on human knowledge as much. You can just gather millions and millions of images and train a computer-vision model. It’s not as obvious how to do that with scientific discovery or policymaking.

You’re applying machine learning to policymaking?

Tavares: That’s an area we’re pursuing. How do you model a city? We’re starting to talk to agencies in New York City. How can we improve the trash problem? How can we reduce homelessness? If we instantiate this policy, what’s going to happen? And the inverse problem: If we want to reduce trash and reduce homelessness, what policies should we instantiate? How should we allocate resources? Could we build multiscale models, which capture different components of the city, in some coherent and cohesive way? And also make it accessible so you can actually help policymakers answer some concrete questions?

Will you be working with the city to answer specific questions about trash pickup, or developing new tools that anyone can use to work on these kinds of problems?

Tavares: We’re starting with particular questions, but to answer those we will require a more general set of capabilities. Can we build a model of a few blocks of New York that are at a level of scale that’s not been done before? That model could then be used to ask a variety of different questions. But just to make sure we’re grounded, we do want to have a particular set of questions.

Witty: One thing that’s especially important is that we want to involve experts and stakeholders, to encode their knowledge, their preferences, their goals.

Tavares: Which is itself quite a hard problem. There’s no massive data set of people’s commonsense knowledge about the urban environment. We’re excited because I think there is a real opportunity to do these two things in tandem—build this foundation of inference but also have an effect immediately.

Will you publish papers?

Witty: Yeah, we’re certainly looking to communicate with the research world. And organizationally, we’re planning on having people work with Basis who are not Basis staff, and often they will be academic researchers with incentives to publish and further their academic careers. One thing I will say is that personally, during my Ph.D., I would often scope projects with the paper as the end goal, and I’m planning on shifting that mind-set to focusing on the work and then afterwards using a paper as a means of communication. But yeah, we don’t want to be hermits in the woods for 20 years, and then come out with this big technology that’s now outdated and totally disconnected from the rest of the world.

Tavares: We are open-source-software-focused, as opposed to the primary output being papers. And within the software focus, we want a unified body of software. We’re trying to build a platform, as opposed to a bunch of different projects.

Could you say more about the organization benefits of having a nonprofit?

Tavares: As a student, your goal is to publish papers and graduate. And that’s only weakly aligned with doing impactful research. We’re working as a team, and our goals are aligned with what we want to do. We’re not unique in that. Look at the papers coming out of DeepMind. They have like 30 authors. I think academia is great for many things, including exploring new ideas. But it is harder, at least in my experience, to build robust technology. It’s not rewarded.

Witty: That’s nonprofit versus academia. From the other side, certainly large tech companies can collaborate in large teams and develop shared infrastructure. But there, there are incentives that maybe get in the way of the work that we want to do as well. The fact that we’re not beholden to make a profit is really freeing.

Will products or services bring income in addition to grants and donations?

Tavares: Hopefully, if we’re successful building what we plan to build, there will be many different domains in which we could. It’s a little bit of a weird new organization. Many things are not certain, and I don’t want to convey things more set in stone or figured out than they are.

The Conversation (0)

Will AI Steal Submarines’ Stealth?

Better detection will make the oceans transparent—and perhaps doom mutually assured destruction

11 min read
A photo of a submarine in the water under a partly cloudy sky.

The Virginia-class fast attack submarine USS Virginia cruises through the Mediterranean in 2010. Back then, it could effectively disappear just by diving.

U.S. Navy

Submarines are valued primarily for their ability to hide. The assurance that submarines would likely survive the first missile strike in a nuclear war and thus be able to respond by launching missiles in a second strike is key to the strategy of deterrence known as mutually assured destruction. Any new technology that might render the oceans effectively transparent, making it trivial to spot lurking submarines, could thus undermine the peace of the world. For nearly a century, naval engineers have striven to develop ever-faster, ever-quieter submarines. But they have worked just as hard at advancing a wide array of radar, sonar, and other technologies designed to detect, target, and eliminate enemy submarines.

The balance seemed to turn with the emergence of nuclear-powered submarines in the early 1960s. In a 2015 study for the Center for Strategic and Budgetary Assessment, Bryan Clark, a naval specialist now at the Hudson Institute, noted that the ability of these boats to remain submerged for long periods of time made them “nearly impossible to find with radar and active sonar.” But even these stealthy submarines produce subtle, very-low-frequency noises that can be picked up from far away by networks of acoustic hydrophone arrays mounted to the seafloor.

Keep Reading ↓Show less