No spike sorting required: C3PO learns interpretable population states straight from raw spike events, from milliseconds to hours
Synopsis
C3PO models neural spikes directly as a marked point process, skipping spike sorting and binning, and uses contrastive prediction over event history to infer a low-dimensional context state. On synthetic data the learned context forms a ring organized by latent phase and recovers tuning curves for 48 phase-tuned neurons; on several published recordings it recovers known variables such as position, head direction and theta phase, and maps population features directly onto single neurons.
Interpretation
An unsupervised method can infer low-dimensional, interpretable population states directly from continuous spike events that have not been spike-sorted, without first assigning spikes to individual neurons and averaging over time bins. Earlier latent-variable methods such as PCA, GPFA, LFADS and CEBRA depend on spike sorting and model binned or smoothed rates, so temporal resolution is capped by the bin width, and nonlinear embeddings make latent directions hard to tie to specific neurons. On synthetic data with 48 phase-tuned neurons plus 16 untuned noise neurons, and with no access to latent phase or neuron identity during training, the first two principal components of the learned context form a ring organized by phase, and the inferred tuning curves closely match the ground truth.
The model compresses each spike waveform into a low-dimensional mark, uses a causal network to infer the current context from event history, and trains on a contrastive objective that scores the true next event against background noise events. The objective derives from contrastive predictive coding and noise contrastive estimation, recast for point processes so that the intractable integral over continuous marks cancels when scores are normalized across the candidate set. On a rat medial entorhinal cortex recording of 460 sorted neurons, context principal components correspond to theta phase, head direction and position, and head-direction tuning strength correlates with each neuron's embedding magnitude in that plane, with a reported coefficient of determination of 0.71.
Perspective
The demonstrated scope covers synthetic spike data, macaque primary somatosensory cortex, rat medial entorhinal cortex, hippocampal CA1 and medial prefrontal recordings, and 51 datasets in the DANDI archive each with at least 50 sorted units and 15 minutes of recording; all models were trained on spiking activity alone, with behavioral and stimulus annotations used only afterward for alignment and grouping.
The authors note that during CA1 sharp-wave ripple replay, C3PO's decoded position representations are less spatially concentrated than a state-space decoder with an explicit continuity model; the slow prefrontal state change cannot be attributed to fatigue, satiation or frustration specifically, so "motivation" is used only as an operational description.
