by Lila Levinsion

Artificial Intelligence (AI) and neuroscience have an extensive and fruitful history of collaboration. Biological neurons are the basis for artificial neural nets (ANNs), a powerful type of AI, and ANNs have, in turn, contributed to our understanding of biological brain circuits. Through major investments from the National Science Foundation, three new centers have been established at UW to help continue this legacy, leveraging bidirectional collaborations between neuroscientists and AI experts to build better AIs and better understand the brain.

AI in everyday life and in neuroscience

AI, a rapidly growing area of research and development as well as a catalyst of social change, is the principle that computers can learn tasks that we traditionally think of as requiring human attention. Many tasks that are trivial to us as humans are extremely difficult for classic computers. For example, if you were given a group of images and asked which ones had cars in them, you would be able to identify cars of different shapes and sizes and from different angles. Until recently, a computer would have struggled to do this – hence the popularity of tasks like this when websites are trying to verify that you are a human, not a robot. You may have recently noticed these tests are getting harder, with blurrier and more ambiguous images. This is because AI has enabled computers to get extremely good at recognizing the content of images, so the “not a robot” bar keeps rising. AI continues to advance extremely quickly and is now being used to create everything from self-driving cars to automated production lines and targeted advertising.

AI is critically important in modern neuroscience research, from helping scientists identify patterns in massive datasets to powering “smart” prosthetics that interface with the nervous system to restore lost functions after an injury such as a stroke. UW’s three new AI institutes each have a unique approach to tap the potential of AI-neuroscience collaborations.

Keeping AIs under control

The AI Institute in Dynamic Systems, headed by CNC faculty member Nathan Kutz, focuses on developing AI technology to control systems with constantly-changing information, such as advanced robotics, brain-computer interfaces (BCIs), and self-driving cars. In the latter example, not only is the technology being deployed in unpredictable environments – such as city streets with distracted human drivers – it is also operating at very high stakes. A self-driving car or a prosthetic arm enabled by a BCI may not be able to tolerate even a low margin of error. These devices require absolute control to guarantee the safety of their users and the people around them.

The Institute will work to build AIs that can operate at this level of control over many domains, including neuroscience. It has collaborators at Harvard and Yale as well as at universities and industry partners throughout the Pacific Northwest. With funding to hire graduate students and postdocs to work with multiple labs, Kutz hopes to enable collaborations that have not otherwise had time to flourish.

Kutz and other Institute faculty will also produce high quality, open-source videos on fundamentals of AI and its applications, as well as on the ethics of AI. They hope to lower barriers to AI education for all interested students, not just those researching with the Institute.

Hardware and software; physicists and neuroscientists

The Accelerated AI Algorithms for Data-Driven Discovery (A3D3) Institute, like the AI Institute in Dynamic Systems, officially began operations at the beginning of October 2021. UW leads nine A3D3 member institutions and is focused on developing AIs across three complex fields – multi-messenger astrophysics, high-energy particle physics, and neuroscience. Amy Orsborn, CNC faculty involved in the latter thrust, is hopeful that spanning disciplines will help her and other researchers develop robust tools to solve common problems.

Orsborn’s lab engineers closed-loop BCIs – hardware that records from the brain, creates some output (such as moving a prosthetic limb) in response to the neural signals, and delivers feedback to the brain about how this output went (such as if the prosthetic limb touched an object in the physical world). AI can be a useful tool for helping a BCI choose an output based on the complex signals it records from the brain. In a real-world application, however, an AI-aided BCI prosthetic would not have an opportunity to learn signals from a robust set of examples – it would need to learn as it went, categorizing the signals in real-time. Not only does this pose challenges to creating successful AI algorithms, it also requires specific hardware.

A3D3, according to Orsborn, will tackle these projects in parallel, “co-designing algorithms and hardware together so they can be implemented in near-real-time.” Astrophysicists, particle physicists, and neuroscientists may work towards different goals, at different scales, with different constraints but, Orsborn says, “when you strip it down to what data you are dealing with, what inferences you want to make, what computations you want to do: they have a lot in common.” To promote collaboration across these fields, A3D3 will hold a series of workshops and events bringing together scientists from each field with AI and hardware developers, and will fund postdocs, graduate students, and summer undergraduate researchers to work across multiple labs.

International AI collaborations

Bringing together labs on an even larger scale, the third recently funded AI institute, the International Network for Bio-Inspired Computing (IN-BIC), will create a “network of networks” to foster international dialogue about AI and neuroscience. The CNC and local partners, including the University of Oregon, the Allen Brain Institute, the University of British Columbia, and the Pacific Institute for Mathematical Sciences, form one hub of this meta-network. The rich, well-established collaborations within this local hub will be connected with similarly-sized groups in Quebec and France that also have robust AI research programs.

IN-BIC is funded by the NSF AcceleNet program, which has the goal of “leverag[ing] research and educational resources to tackle grand research challenges that require significant coordinated international efforts.” CNC co-director Adrienne Fairhall is excited for the opportunity and motivation to interact with international research through workshops and student exchanges. A workshop on AI ethics in Paris is already planned for 2022 and, Fairhall says, the group is currently in talks with the French Attaché for Science and Technology about hosting a workshop in Seattle as well. As its name implies, IN-BIC’s research focus will be on how what is known about the brain can be applied to modern computational problems and vis versa. Fairhall says that primary research questions will range from “how can we use AI…to accelerate the way we do neuroscience?” to “what is different between a biological brain and an algorithm … [and] what’s special about neurons as computational devices?” IN-BIC has the potential not only to elucidate and expand on the already rich bidirectional interactions between AI and neuroscience research, but also to foster a global community of scientists.

UW and the future of AI

Each of these institutes alone presents exciting new opportunities, but together, they represent a huge investment in UW’s AI focus. “There are very few places in the world that can compare,” says Kutz. “When we look back at this century, I think Seattle will have played a huge role in whatever the world looks like in the year 2100…we [will have] educated the talent pool that put those things out there.”