Generalization with simultaneously recorded populations

[vishal.neuro]

Dear Neural Decoding Toolbox developers,
We are trying to use your toolbox to check the generalization abilities of a simultaneously recorded population. In order to avoid a biased estimation of the generalization error, it is important to make sure training and test set do not share data that were recorded at the same time, so we used the following option
create_simultaneously_recorded_populations=1 that we assume makes sure the datasets are split properly.
While this approach runs without trouble for a “basic_ds” object, setting this parameter for a “generalization_ds: object, leads to the error
Error using basic_DS/get_data (line 421) problem, all simultaneously recorded neurons should have the same labels

We are wondering:
– Whether we are correct regarding the role of “create_simultaneously_recorded_populations=1 “
– Whether this is a bug, or we are simply supposed to use “create_simultaneously_recorded_populations=0“ in the case of generalization.
– We assume using “pseudo populations” should not lead to the same validation procedure and results, compared to populations, because the former would draw both units and trials at random, instead of drawing all available units from the same trial for a single data point of the train/test set. Is this reasoning correct? Is drawing all available units from the same trial implemented in the toolbox?

Many thanks

[Author]

Posts