The MARBLE method is a fully unsupervised representation learning approach to obtain interpretable latent representations of neural dynamics. More generally, it introduces a statistical learning paradigm for non-linear dynamical systems based on a decomposition of the dynamical attractor into local flow fields. MARBLE representations achieve state-of-the-art decoding accuracy in neural dynamics and allow comparing computations across biological and artificial neural networks.

Develops a discrete geometric measure, the dynamical Ollivier-Ricci curvature, to discover a fundamental link between the information flows over the network and its connectivity. Using this geometric theory, it describes the thermodynamic detection limit of network clusters. This geometric theory leads to a practically useful algorithm for computing information propagation in networks.

Our method demonstrates the possibility of monocular 3D pose estimation (i.e., using one camera only) in freely behaving laboratory settings with few training poses, hardware limitations and occluded body parts. It opens the door to 3D kinematic analyses of laboratory animals in biomechanical and robotic studies, where data was limited to 2D poses.