Why We Need Computation To Understand the Brain

“At the time, there were only a handful of initiatives focusing on computational neuroscience work,” says Alyssa Picchini Schaffer, vice president and senior scientific officer of the Simons Foundation’s neuroscience collaborations. “There was this realization that deeply grounding the theory in biology and, on the flip side, having experimental neuroscientists really dig in and understand the computational approaches, that was the special sauce.”

It was quickly apparent that uncovering the patterns of neural activity that underlie cognition requires studying many regions of the brain at once, and that meant large-scale science, in terms of both the amount of data generated and the number of investigators with different areas of expertise involved. Perhaps most emblematic of the SCGB’s big, collaborative science ethos is the International Brain Laboratory (IBL), an SCGB-funded consortium of 22 neuroscience labs that came together in 2017 to map neural activity across the entire mouse brain as animals engaged in a single task.

Their brainwide map of neural activity from 300,000 neurons, finished in 2023, required new computational tools for storage and analysis, developed by a dedicated team of 12 engineers who support the IBL. This large dataset of neural activity, a first of its kind, has been made openly available for exploration and insight generation to the broader neuroscience community. Computational neuroscientists working in the IBL have also developed approaches based on their data that could have broad applications across systems neuroscience.

The thorny problem of big data analysis has reached many corners of neuroscience. Byron Yu, a neuroscience and engineering professor at Carnegie Mellon University and an SCGB investigator, has been working for more than 10 years on a collaborative effort to understand how different parts of the brain interact to enable vision and interpretation of what we see. That effort requires tracking neural activity from hundreds of neurons across those different parts of the brain in one experiment. In traditional neuroscience studies that measured the activity of only a few neurons at a time, scientists typically studied the activity of each neuron individually or that of pairs of neurons. But for these large, more complex datasets, Yu and his colleagues needed to apply multivariate statistical methods (a process known as ‘dimensionality reduction’), which make it possible to study how larger numbers of neurons coordinate their activity together. As datasets get larger across systems neuroscience, this statistical approach is becoming more and more common.

Yu specializes in data analysis and statistics, while his collaborators — Adam Kohn, chair of neuroscience at Albert Einstein College of Medicine, and Christian Machens, a neuroscientist at the Champalimaud Foundation — lead the project’s laboratory experiments and modeling efforts, respectively.

“Experimentation, modeling and analysis are like three corners of a triangle. Each one has its own value, but when you put the three together, that’s where the power comes from,” Yu says. “The meeting point is really interesting, where the data meet the model and we can compare and contrast.”

The SCGB has also invested in training the next generation of neuroscientists. The collaboration funds fellowships for postdoctoral fellows and those transitioning to academic independence. Laura Driscoll, now a scientist at the Allen Institute for Neural Dynamics, received one of those fellowships to support her research and transition, building on her postdoctoral work with David Sussillo and the late Krishna Shenoy at Stanford University.

One clear problem in machine learning is the difference between our brains and artificial neural networks in how they adapt to variations. Our brains are very good at learning new skills by making inferences from skills we already know, but computers are less readily adaptable in this way. Driscoll’s project studied how an artificial neural network trained to complete multiple tasks used sections of its activity patterns repetitively to accomplish those tasks, the way a dancer might use the same basic dance moves in different choreographed routines. Understanding how an artificial network is able to complete multiple tasks that it has been trained on is the first step toward building networks that can flexibly learn new skills.

“The work I was doing wouldn’t have been possible if it weren’t for the Simons Foundation, because there weren’t many other funding mechanisms for people who are doing simulations in artificial neural networks,” Driscoll says. “I feel a lot of gratitude toward the SCGB for being able to have the career that I’ve had so far.”

The SCGB brought together all its investigators and fellows for a regular annual meeting, and several SCGB scientists say that meeting was the highlight of their year — and even helped change the course of their careers.

“I don’t think I’ve ever been in a group of such distinguished scientists as at the yearly meeting,” says Sussillo. “In terms of what they set out to do, it worked. Those ways of thinking have now become mainstream, thanks to the Simons Foundation. What an honor to be a part of it.”