AI researchers have been turning to the brain for inspiration since the emergence of the field. Although AI agents have achieved super-human performance in some tasks, there are still many gaps. Our brains are very versatile; humans can solve many problems without having to learn each and every thing from the beginning, all while using very little energy. And how about creativity? While there is tremendous progress at closing the gaps, our limited understanding of the brain hinders advances. The brain remains mostly a black box.
Neuroscience is the science of reverse engineering the function of the brain. For example, neuroscientists use imaging techniques to study how memories are formed, how attention is controlled, and how we communicate. In the Brain-Inspired Computing Lab, we partner with academic researchers who work at the intersection of neuroscience and AI. Our goal is to develop neuroscientifically-inspired learning algorithms and technologies that make new computing possible. In support of our goal, we also contribute to neuroscience research which feeds back to AI results.
Our first open source contribution to the neuroscience community is the Brain Imaging Analysis Kit (BrainIAK), a software library for analyzing neuroimaging data. BrainIAK is the result of our close collaboration with the Princeton Neuroscience Institute. It is primarily aimed at functional magnetic resonance imaging (fMRI), a non-invasive technology often preferred for cognitive neuroscience research because of the high-resolution images of full-brain neural activity it produces. Each fMRI scan produces a 3D image composed of up to one million volumetric elements, or voxels. Scanning with sampling intervals of one second leads to tens of gigabytes of data collected per subject per study. Large neuroscience studies can include thousands of people, resulting in staggering amounts of data. This makes applying advanced imaging analysis methods computationally challenging and is the reason BrainIAK was built to be scalable and computationally efficient.
BrainIAK is easy to install via Python* packages, Conda*, and even Docker* container images. BrainIAK can run unmodified on regular laptops as well as in the cloud and on high-performance computing clusters. There are multiple ways for users to interact with BrainIAK: Jupyter* notebooks, which have easy-to-use browser-based interfaces, regular Python applications, or batch jobs for clusters. Behind a simple Python interface, BrainIAK offers parallelization capabilities via MPI and OpenMP. It uses open standards such as NIfTI and BIDS for input/output. It also offers several frameworks and tools that can be used by many of the analysis methods, such a fMRI data simulator or hyperparameter optimization for machine learning models.
The most important part of BrainIAK consists of the advanced fMRI analysis methods. Almost all of these are new methods that were contributed by their authors, often at the same time that they were published in the scientific articles that first introduced them. We added new multi-subject analysis which pools together data from thousands of subjects, increasing the statistical power of the experiments. We introduced correlation-based analysis methods that can capture the distributed nature of neural representations. With the Bayesian representational similarity analysis algorithm, we added a new way to decrease the noise in fMRI data, which is a limiting factor for many analyses. And we contributed event structure detection, which is based on a promising hypothesis about brain activity structure, which can detect when a subject is remembering the same episode that they previously listened to in a recorded story.
We have taken a number of steps to introduce BrainIAK to the neuroscience community. We organized hackathons at Princeton, the University of Texas at Austin, Yale, and Virginia Tech. We also participated in code sprints at Stanford for the promotion of a common way of organizing and sharing fMRI data. We presented BrainIAK-focused posters at the Society for Neuroscience (SfN) and Organization for Human Brain Mapping (OHBM) annual meetings. Our latest poster at SfN, presented this month, focuses on educational materials developed around BrainIAK for university courses on advanced fMRI analysis at Princeton and Yale. Finally, we organized a workshop about high-performance computing techniques at the largest conference in the world for neurofeedback researchers.
BrainIAK helps neuroscientists understand the brain. In turn, this new understanding provides development avenues for AI. For example, we have discovered multi-tasking limitations in artificial neural networks by observing how the brain develops new skills. We invite researchers to contribute to this community effort and take advantage of BrainIAK in their own work. We are excited to see the next breakthroughs powered by BrainIAK and where they will lead AI.