Fig. 5: Interpretable representation and cross-animal decoding of neural activity in rat hippocampus during linear maze navigation.
From: MARBLE: interpretable representations of neural population dynamics using geometric deep learning
a, Experimental setup of a rat navigating a linear maze with tracked position and direction of motion (top). Raster plot showing spiking activity in 120 neurons in a single session (bottom). Rat image credit: [designer_an]/stock.adobe.com. b, Comparison of latent representation (E = 3) of unsupervised MARBLE against self-supervised (time-only labels) and supervised (time, position and direction labels) CEBRA. Color shading is defined in a. c, Time traces of linearly decoded animal position within rat 1 (E = 32, default settings from CEBRA decoding notebook examples). d, Decoding accuracy within the same animal. Two-sided Wilcoxon tests (paired samples), ****P < 1 × 10−4. Horizontal and vertical bars show mean and 1 × s.d., respectively (n = 2,000). e, Cross-animal consistency as measured by R2 of linear fit between the optimally aligned 3D latent representations of a source animal to a target animal.
