JEDI: Jointly Embedded Inference of Neural Dynamics

A research team from Université de Montréal and Mila AI institute has introduced JEDI (Jointly Embedded Inference of Neural Dynamics), a breakthrough hierarchical model that addresses a fundamental challenge in neuroscience: understanding how brains flexibly adapt to different tasks using the same neural circuitry. Traditional recurrent neural networks (RNNs) trained on neural data have been limited to single-task domains and struggle with generalization across behavioral conditions. JEDI overcomes this by learning a shared embedding space over RNN weights, allowing it to capture neural dynamics across multiple tasks and contexts within a single unified framework.

The model demonstrates remarkable capabilities in both simulated and real-world applications. Using simulated RNN datasets, JEDI accurately learns robust, generalizable, condition-specific embeddings that scale to arbitrarily large and complex datasets. When reverse-engineered, the model successfully recovers ground truth fixed point structures and reveals key features in the eigenspectra of underlying neural dynamics. Most impressively, the researchers applied JEDI to actual motor cortex recordings from monkeys performing reaching tasks, where it extracted mechanistic insights into the neural dynamics of motor control. This represents a significant advancement in reverse-engineering brain function from experimental recordings alone.

The JEDI framework's ability to jointly learn contextual embeddings and recurrent weights provides neuroscientists with a powerful new tool for understanding how neural populations achieve behavioral flexibility. By uncovering shared structures across different conditions, the model offers a more comprehensive view of brain dynamics than previous approaches. This work bridges the gap between theoretical neuroscience and practical AI applications, potentially accelerating our understanding of neural computation and inspiring more flexible AI architectures.