Fig. 3: Hourly time prediction by polling randomly selected cohorts of SCN neurons.

a Circular evolution trajectory of a 2D representation of population-level Ca²⁺ activities. Axes correspond to the first two principal components, PC1 and PC2. Different time points are represented in different colors, and consecutive time points are connected by dotted lines. b Scheme of time-predictor based on SCN Ca²⁺ signals. Conv and FC refer to the convolutional and fully connected layer; N, T and D denote the dimensions of neurons, the input Ca²⁺ time series of each neuron and its features extracted, respectively. c Accuracy and loss curves during training process. d Prediction accuracy curves, showing results from six SCN slices. Data are shown as mean ± SEM (n = 5000 trials). Lower dashed line represents the chance level. e Visualization of CNN’s high-dimensional features in a 2D space via t-SNE. From left to right: the number of neurons is 1, 100, 300, and 900, respectively.