A critical initialization for biological neural networks

Key Points:

• The study conducted 26 neural recordings in mice, targeting cortical and hippocampal regions, using genetically encoded calcium indicators and Neuropixels probes, following approved ethical protocols at HHMI Janelia.
• Neural activity was recorded using custom-built two-photon mesoscopes and processed with Suite2p for motion correction, ROI detection, neuropil correction, and spike deconvolution, with advanced methods improving spike inference accuracy.
• Data analysis included normalization, clustering, eigenspectrum estimation via covariance between neuron subsets, and estimation of principal component timescales and rotational dynamics using Dynamic Mode Decomposition (DMD) with ridge regression.
• Theoretical modeling employed linear dynamical systems with stochastic inputs, deriving power-law decay of covariance eigenvalues and relating eigenvalues to neural timescales; simulations of 10,000 neurons validated these analyses and explored connectivity structures.
• Additional analyses investigated local correlation structures, decoding of inputs from network activity, and comparisons with rodent and primate neural data, providing insights into the dynamics and connectivity underlying spontaneous and task-evoked neural activity.