Forum Replies Created
-
Posts
-
The “weights” used will depend on the classifier used (some classifiers have weights others do not). I don’t think there is a function built in that returns the weights particularly since they change every time the classifier is trained.
There are two options if you are interested in questions of what I was calling “information sparsity” (or what I called “compactness” in my 2008 J. Neurophysc paper). The first is to write your own custom code to capture this information from the classifier which might not be that easy. The second way (which is much easier) is to run the analysis using only the k most selective neurons for a range of k values (e.g., run the analysis for k = 2, k = 4, k = 8, etc.). This can be done using the feature preprocessor select_or_exclude_top_k_features_FP. If the classifier works just as well using only 8 neurons as it does using 100’s of neurons then this tells your there is a information rich sparse subset of neurons that contains all the information that is available in the whole population (check out figure 4 of my 2008 J. Neurophys paper and figure S2 of my 2012 PNAS paper).
I hope that helps!
The index used when you set ds.sites_to_use is the index that the data is stored in the binned_data cell array. This should be the same order as that the raster files are loaded (based on the create_binned_data_from_raster_data function). You do not need to add a filed in raster_site_info to get this working – although having meta information in raster_site_info might help to make the correspondence between the index of each neuron and other information about the neuron when looking at the binned_data.
Hi Simone,
Sorry for the slow reply. Yes, the toolbox can work with LFP/EEG/MEG. We have had success decoding filtered MEG signals and using the regular create_binned_data_from_raster_data function (see The dynamics of invariant object recognition in the human visual system). This might also work for LFP and EEG signals, although it is possible that you might get better results writing your own binning function that creates features that are better suited for these different signals. Overall which types of features work best for decoding on any type of signal (including neural spiking activity) is an open question, so there is potentially room for improvement (and for learning more about neural coding).
Hope that helps!
Ethan