While optical methods, genetically encoded fluorescence indicators, and optogenetics already enable fast readout and control of large neuronal populations using light, the lack of corresponding advances in computational algorithms have slowed progress. The fundamental challenge is to reliably extract spikes (neural activity) from fluorescence imaging frames at speeds surpassing the indicator dynamics. To meet these challenges, we devised a set of new algorithms that exploit tensor-based computing on accelerated hardware. We provide optimized motion correction, source extraction and spike detection operations, which for the first time operate at speeds comparable with brain internal communication. We evaluate these algorithms on ground truth data and large datasets, demonstrating reliable and scalable performance. This provides the computational substrates required to interface precisely large neuronal populations and machines in real-time, enabling new applications in neuroprosthetics, brain machine interfaces, and experimental neuroscience.
