OPTIMAL TRANSPORT THEORY TO EXTRACT SPIKING PATTERNS

Whether cortical neurons encode information through firing rates or precise spike timing remains a central question in neuroscience. Although spike timing can theoretically convey more information than firing rates, its utility is often challenged by neural variability or synaptic noise. For these reasons, precisely estimating spike timing can be a challenging task. Here, we investigate the use of the Earth Mover’s Distance (EMD), a metric from optimal transport theory, to extract spatiotemporal spiking patterns with a single-layer autoencoder (AE). We compare AEs trained with the EMD or the mean squared error (MSE) on synthetic spike trains and on large-scale neural recordings from the Allen Institute. Using synthetic data, we demonstrate that EMD-based training is markedly more robust to temporal jitter and time warping, whereas the MSE is less sensitive to additive noise and probabilistic participation of the neurons. When many samples are available, MSE-trained AEs better capture temporal sequences along with spike-timing precision. However, EMD-trained AEs reliably estimate the averaged spike timings with relatively few samples, which is often the case in real-world experimental conditions. Applied to mouse visual cortex recordings, the EMD enables the extraction of temporal sequences that discriminate visual stimuli more effectively than the MSE, and capture different information than what can be achieved using averaged firing rates alone. Altogether, these results demonstrate that optimal transport theory provides an efficient tool for uncovering spatiotemporal structure in spiking activity.