Advancing neuroscience research with AI
To understand the brain, we have to measure behavior and neural activity precisely, and then relate the two. In this talk I will show how my lab builds open-source machine learning tools for both. DeepLabCut uses deep learning and transfer learning to do markerless pose estimation, tracking animal and human movement from ordinary video with little labeled data. It now supports multiple animals, real-time use, and pretrained “SuperAnimal” foundation models that work across species and settings. CEBRA uses contrastive learning to find consistent latent embeddings that link behavior and neural recordings. These embeddings can be compared across sessions and animals and used for decoding, including reconstructing natural video from mouse visual cortex activity. In closing, I will discuss how tools like these, and openness in science, can help move us toward a computational understanding of adaptive behavior.
Mackenzie leads the Mathis Lab at EPFL and created DeepLabCut, open source pose estimation software now used across thousands of biology labs. A model case for how open tools reshape an entire research field.