Fig. 4. A low-rank model captures task-relevant LIP activity during the DMS and DMC tasks.

A We determined the dimensionality of the population activity by estimating model fitness while increasing the number of components. Each point shows the mean cross-validated log-likelihood per trial of the GMLM relative to a baseline model without any stimulus terms (i.e., the “rank 0” model) for two monkeys during the category task. The log-likelihood shows the average log probability of observing the spike train from a withheld trial from one neuron given the model relative to the log probability of that observation under the null model without stimulus-dependent terms. In a Gaussian model, the log-likelihood is proportional to the squared error and it is related to the variance explained. Here, we instead use the Poisson likelihood which is more appropriate for quantifying spike count observations^37,80. The traces show the GMLM with different amounts of category or direction information included. The dashed lines show the corresponding GLM fits (the full-rank model). The GMLM accounts for most of the log-likelihood with a small number of components. B Number of stimulus components (rank) selected for the GMLM for each LIP population to account for 90% of the log-likelihood. The number of stimulus parameters in the low-rank GMLM is compared to total parameters in the equivalent single-cell GLM fits for each population.