Supplementary Figure 5: Four temporal basis components are sufficient to describe the receptive fields of all neurons in the population.

To find shared temporal basis components across neurons, we can simply stack the VWAs of all of the neurons together (to form a 100 by (72*N) matrix where N=number of neurons) and run singular value decomposition on this large matrix. The first few temporal components will be selected to allow for optimal squared error reconstruction of all of the VWAs. From left to right, the plots depict the predictive performance of various reduced rank models (reconstructed from 1 to 4 shared temporal basis components) against full rank models (100 temporal basis components). The rank refers to the rank of the RF matrices with temporal basis components shared across neurons. Critically, the temporal basis components are the same for all neurons, so the low number of components required to match full rank performance demonstrates that there is shared structure in the temporal components of the VWAs across all neurons. By rank 3, the low rank model predictive performance is approximately equivalent to the full rank model, demonstrating no more than 3 or 4 temporal basis components are required to capture the RFs of any neuron in the population.