Extended Data Fig. 5: 40 ms CNNs are resilient to oscillatory substates.

A, Top- Broadband trace of exemplary activity during active wake in MOp over several seconds. Blue box shows the width of an individual input data used by the 40 ms CNN to predict state. Middle- Raster of MOp spiking. Bottom- Stacked barplot of CNN prediction probabilities across the three states every 1/15 s. B, Exemplary data from MOp during quiet wake. C, Cross-correlogram between spindles (S), ripples (R), and flickers (F) in Animal 5. A strong central peak in the cross-correlogram is observed between spindles and ripples consistent with prior work (Siapas & Wilson⁷⁹). No substantial positive correlation is observed with flickers by spindles or ripples. D) Cross-correlogram between OFF-states and flickers of various states (for instance ->W includes NREM-to-wake, and REM-to-wake). A major central correlation trough is observed in flickers to wake, meaning flickering is reduced during sleep OFF states. E) Percent of substates which coincide with errors in model classification in 1 s CNNs. Stacked black lines show the intermediate quartiles with all points as swarm scatter colored by region. Cortical OFF states and sleep spindles were detected in all cortical regions in two animals (n = 10). Ripples were detected in all recordings of CA1 hippocampus (n = 7). As a negative control, for each substate, an equal number of randomly selected timestamps were selected and evaluated. No significant differences (p > 0.05) were found between this negative control and any of the three substates.