Fig 1.
(A) Examples of spiking activity recordings in the CRCNS fcx-1 dataset in the WAKE state. Left: spike train raster of a random subset of excitatory cells (red) and inhibitory cells (blue). Right: examples of ISI series produced from spike train chunks of inhibitory/excitatory cells in the fcx-1 dataset. (B) Interspike interval value distribution histograms generated from the aggregated spike trains of retinal ganglion cells in response to a “white noise checkerboard” visual stimulus (red) and a “randomly moving bar” stimulus (green). (C) Interspike interval value distribution histograms generated from the aggregated PFC spike trains (fcx-1 dataset) corresponding to the WAKE (blue) or SLEEP (yellow) state of the rat.
Table 1.
Spike train classification results for the retinal neuron activity dataset for nearest-neighbor models with different distance metrics.
The task is defined as binary classification of the stimulus type (“white noise checkerboard” or “randomly moving bar”).
Fig 2.
Spike train classification metric values (top panel—accuracy, bottom panel—AUC-ROC) for the retinal neuron activity dataset on a range of models.
The task is defined as binary classification of the stimulus type (“white noise checkerboard” or “randomly moving bar”), with the test set balanced in class distribution by undersampling (that is, accuracy = 0.5 corresponds to chance level). Models are ranked in ascending order of the median metric value. The “simple baseline” model tag corresponds to spike trains encoded with 6 basic distribution statistics, the “raw” tag implies that the model has been directly trained on ISI time-series data without feature extraction. The “tsfresh” tag corresponds to encoding with the full set of time-series features. “ISIe” stands for interspike-interval encoding of the spike train, “SCe” stands for spike-count encoding. “ISIe + SPe” means that feature vectors corresponding to both types of encoding are concatenated. InceptionTimePlus, FCNPlus, and ResNetPlus refer to implementations in the PyTorch-based tsai package.
Table 2.
Balanced test set accuracy values on the retina dataset with different classifiers trained on tsfresh feature representations obtained from the (i) interspike interval encoding of the spike trains, (ii) the spike count encoding of the spike trains, (iii) combined interspike interval + spike count encoding.
Table 3.
Classifier performance degradation in case when no preprocessing for ISI-encoded data is used compared to our standard preprocessing pipeline.
Cohen’s kappa score for a range of models in the retinal neuron activity task is reported.
Table 4.
Spike train classification metric values (for imbalance-robust metrics) for the retinal neuron activity dataset on a range of models.
The “simple baseline” model tag corresponds to spike trains encoded with 6 basic distribution statistics, the “raw” tag implies that the model has been directly trained on ISI time-series data without feature extraction. The “tsfresh” tag corresponds to encoding with the full set of time-series features. “ISIe” stands for interspike-interval encoding of the spike train, “SCe” stands for spike-count encoding. “ISIe + SPe” means that feature vectors corresponding to both types of encoding are concatenated. InceptionTimePlus, FCNPlus, ResNetPlus and XceptionTimePlus and refer to implementations in the PyTorch-based tsai package.
Table 5.
Spike train classification accuracy values for different datasets on a range of models.
The reported accuracy is measured on balanced test sets to mitigate class imbalance, median value and standard deviation in percent are shown. Model names correspond to the same ones from Table 2.
Fig 3.
Classification accuracy for the Allen cell types VIP/SST interneuron classification task in the case of multiple randomly sampled same-class spike train chunks per prediction (with prediction done via majority voting).
The model trained in these trials is a random forest classifier on the full set of tsfresh features. The boxplots reflect the median accuracy and the variance between different train/test splits as done in the main text for the fcx-1 data set. The red crosses correspond to the theoretical estimate under the assumption of independently sampled spike train chunks.
Fig 4.
Spike train feature embeddings for WAKE (points marked red) vs. SLEEP (points marked blue) activity states of the neural circuit.
Two-dimensional embeddings of the (20-dimensional) selected-tsfresh-feature space using (A) unsupervised UMAP and (B) supervised UMAP embedding algorithms for spike trains corresponding to WAKE vs. SLEEP activity states.
Table 6.
Test set AUC-ROC values for the unsupervised temporal structure recognition tasks for different base spiking datasets and different transforms.
A random forest classifier model (see S1 Text for the hyperparameter values used) was used in all of the above experiments.