Audrey Woods, MIT CSAIL Alliances | August 19, 2026
For years, the grand ambition of neuroscience-informed computer scientists was to understand the brain well enough to build better AI. In 2017, Google DeepMind co-founder Demis Hassabis wrote a widely read argument that a “better understanding of biological brains could play a vital role in building intelligent machines.” Then large language models arrived and proved they didn’t need a deep understanding of the human brain to respond with human-like language. “[Hassabis] wrote that right before GPT. Since then, he—and others in the field—have backed off on that stance,” says MIT CSAIL graduate student Christopher Wang.
But now, Wang is exploring the opposite route: trying to take what has worked in artificial neural networks and use that to understand the human brain. He believes the two directions may not stay separate for long and could, in fact, “end up coming full circle.”
Working in the InfoLab with Principal Research Scientist Boris Katz, Wang has spent his PhD building models and datasets that map how language moves through the brain in space and time. By learning how to best collect and apply neuroscience data, Wang hopes we will one day understand the human brain so deeply that we might even be able to replicate it.
FROM LINGUISTICS TO NEUROSCIENCE: BUILDING DATASETS FOR THE BRAIN
Wang did not set out to study the brain. “When I wrote my PhD application, I did it for natural language processing. I was really into linguistics as an undergrad. I wrote my essays all about language.” He submitted those essays before ChatGPT arrived and, as he puts it, changed “everything in natural language processing forever.”
The first project he joined at MIT was a collaboration with Boston Children’s Hospital to study language processing in the brain by correlating neural activity with what patients saw on screen while watching a film. Wang was hooked. “I was interested in language, so I was really down to ask, okay, how does language work in the brain? Is it the same way that it works in machines? And all of that became: how does the brain work, period? That’s how I ended up in the neuroscience space.”
Wang’s work is organized in the stages of a machine learning pipeline: data, modeling, and evaluation. High quality data of the brain is difficult to come by, which is why Wang and his InfoLab colleagues rely on a cohort of epilepsy patients at Boston Children’s Hospital who, during a week-long protocol for evaluating seizures, have electrodes implanted in their brains. While there, some of these patients volunteer for language studies. Historically it’s been difficult to collect such data because the standard protocol of patients reading sentences off a screen one word at a time is tedious, especially for children. “If you’re a kid and you’re in the hospital and you have wires sticking out of your head, you really don’t want to sit through a long, boring language experiment.” Wang’s team came up with a more practical way to link brain activity with language: movies. Using well-known, kid-friendly feature films with scripts that can be easily mapped to specific time and brain activity, they are able to gather high-quality data about what the brain is doing as it interprets language.
This work has led to the Brain Treebank, a dataset of intracranial recordings from ten subjects who together watched 26 films, roughly 43 hours, while probes captured their neural activity. Because the movie scripts are known, every one of the 223,000 spoken words can be aligned to the exact moment the brain heard it, then labeled by part of speech and parsed for grammar. This data volume allows Wang and his fellow researchers to “ask all sorts of interesting questions.”
But how to answer those questions? To address the modeling step of Wang’s work, he created BrainBERT, a system that learns to represent short periods of brain activity and group them by similarity, producing a block-by-block picture of what the brain is doing over time. Using such models, his group can train a decoder to look at neural activity and tell whether a person is hearing a noun or a verb, then repeat that across every recording site to build a map of where different pieces of language processing happen. The results mostly confirm what neuroscientists already suspected—“we needed to ground what we found in the current literature”—but BrainBERT also revealed a new level of resolution. “What we find is a greater spatial and temporal precision than what was known previously.” This lets researchers trace the order in which the brain handles things like pitch, loudness, part of speech, and meaning as a sentence unfolds.
Wang wants other researchers to be able to run the same evaluations his group does, and to exceed them. To that end, he has been publicizing Neuroprobe, a benchmark that turns the recorded hospital data into a battery of standardized decoding tasks with a public leaderboard. “A lot of the time a field is waiting for a good benchmark before it starts taking off,” so he aims to build exactly that benchmark to accelerate a field he’s passionate about. It is still a preprint, but some submissions are already outscoring the models his team built. “This is exactly what we want. This is an area where we really want to be beaten.”
DIGITAL TWINS AND THE LAST GREAT FRONTIER
The holy grail of this work, Wang says, is to eventually be able to create a full, accurate digital twin of the human brain. “Once you have that, you can do so many things on the clinical side. You can start simulating the outcome of different treatments. You can start simulating different processing under different settings. We're getting to the point where we have enough data collected from among enough different modalities that building these digital twins is not such a crazy ask anymore.” In the not too distant future, we might even be able to fully simulate the brain’s activity and create a new kind of artificial intelligence. “It'll be such a cool GPT moment as well, because it won't be the old dream. The old dream was to understand the brain so well that we can write down with pencil and paper what it's doing, and then get a machine to do the same thing. The current approach is to train from scratch a model on inputs and outputs. If people are able to get simulations of a high enough quality, I would say it would be a completely new pathway to intelligence,” which Wang “takes for granted would be conscious.”
While fully simulated brains are probably many years away, Wang is excited about the near-term impact their techniques might have on healthcare pipelines and clinical practices. Right now, the standard of care in epilepsy monitoring—where their current data comes from—requires neurologists to comb through terabytes of brain recordings from the week of a patient’s stay. Pinning down exactly where seizures begin and end, the first step in charting how they spread, requires enormous expert time and effort. Wang believes models like BrainBERT could scan through the data much faster, roughing out a coarse version of the timeline so that the specialist can get to the relevant moments more efficiently and spend more time on them. In fact, a new collaboration with Massachusetts General Hospital aims to turn the group’s brain models into a deployable, time-saving clinical tool. Further out, better models of the brain could improve brain-computer interfaces like the ones Neuralink and Meta are pursuing. “If their bet ends up panning out, that’ll be the big thing for the next ten years.”
Wang’s personal motivation stems from the fact that the brain is the “last thing we’ve not understood well.” We have computers that can play chess, write code, and produce coherent language, all the while skipping past a concrete understanding of how the brain does those things. “We’ve gotten to intelligent systems, but we’ve completely bypassed understanding the principles that underlie human language processing. It’s incredibly cool that there’s this thing we all have access to, it’s been with us from the beginning, and no one understands it yet.” Wang says, “the chance to work on that is, I think, amazing.”
Learn more about Wang’s work on his website, Google Scholar page, or CSAIL page.
