Figure 1. Functional Encoding Unit pipeline and validation.

A. We combined a state-space approach for modeling neural firing with an unsupervised approach for clustering neurons based on their stimulus-response function. We simulated neural activity in a population of 50 noisy neurons belonging to five ground truth clusters of possible responses to a stimulus: excited-sustained, excited-phasic, inhibited-sustained, inhibited-phasic, and nonresponsive. To evaluate the accuracy of our approach for correctly identifying ensembles and making inferences about their parameters, we analyzed this simulated data using our Functional Encoding Unit (FEU) pipeline.

B. The Functional Encoding Dictionary (FED) accurately clusters noisy neurons into the five ground truth ensembles: excited-sustained (FEU1), inhibited-sustained (FEU2), inhibited-phasic (FEU3), excited-phasic (FEU4), nonresponsive (FEU5) (48/50 = 96%). Each FEU has two parameters that define its function: the Jump (J) parameter and the Phasicity (P) parameter. J describes how much neuronal responses within an FEU are modulated by a given stimulus. A positive value indicates excitation, while a negative value indicates inhibition. A value close to zero (between −0.1 and 0.1) indicates no response. P describes how phasic or sustained the FEU response is to a given stimulus. A small value of P (< 5) indicates a sustained response, and larger values of P indicate a more phasic response. The FED provides parameter values for J and P that accurately model the ground truth relationships between FEUs.

C. 2-D graphical representation of FED. FEUs have centroids corresponding to their value of J and P parameters; the size of the FEU represents the relative number of neurons in each FEU. Callout to rasters for each FEU made from the individual rasters of all neurons in that FEU.