Fig. 4: Interpretable representation and decoding of neural activity during arm-reaching.
From: MARBLE: interpretable representations of neural population dynamics using geometric deep learning
a, Ground truth hand trajectories of a macaque in seven reach conditions. Monkey image adapted by Andrea Colins Rodriguez from https://www.scidraw.io/drawing/445, CCBY 4.0. b, Single-trial spike trains in the premotor cortex for three reach conditions (24 recording channels within each color). The shaded area shows the analyzed traces after the GO cue. c, Firing rate trajectories for a reach condition (up) PCA-embedded in three dimensions for visualization. d, Vector field obtained from firing rate trajectories. e, Latent representations of neural data across conditions in a single session. CEBRA-behavior was used with reach conditions as labels. The MARBLE representation reveals as an emergent property the latent global geometric arrangement (circular and temporal order) spanning all reaches reflecting physical space. f, Linear decoding of hand trajectories from latent representations. g, Decoding accuracy measured by R2 between ground truth and decoded trajectories across all sessions for the final position (left) and instantaneous velocity (right). Two-sided Wilcoxon tests (paired samples), **P < 1 × 10−2; ***P < 1 × 10−3; ****P < 1 × 10−4; NS, not significant. Horizontal and vertical bars show mean and 1 × s.d., respectively (n = 43).
