Problem with labels

1

Hi Martin (and all TDT experts),

I recently started to use TDT and I’m basically moving my data from CoSMoMVPA to TDT. This wasn’t a problem until I got this error:

Error using ==
Matrix dimensions must agree.

Error in decoding_transform_results (line 98)
output = 100 * (1/size(predicted_labels,1)) * sum(predicted_labels == true_labels); % calculate mean (faster than Matlab function)

Error in decoding_generate_output (line 35)
output = decoding_transform_results(curr_output,decoding_out,chancelevel,cfg,data);

Error in decoding (line 568)
results = decoding_generate_output(cfg,results,decoding_out,i_decoding,curr_decoding,current_data);

Error in my_script
[results, cfg] = decoding(cfg);

I’m running an ROI analysis, lda classifier and using “make_design_cv”. I have my data preprocessed in SPM, but I want to use the labels in a different way so I manually defined chunks, names and labels using “cfg.files”. This was not a problem until I had to change the beta files (these betas were created using cosmo functions). I had to use different betas in a second analysis (but still using the same logic as before), and I got this error. The betas were created in the exact same way and the struct is the same. I don’t understand why are the predicted labels and true labels are not matching.

Any ideas about this problem?

Thank you in advance!

Lénia

2

Hi Lénia,

Ouch, that sounds annoying! Could you please try another classifier first, e.g. 'libsvm', to rule out this is an issue with our LDA implementation? I think that might be the culprit.

If it works with libsvm, then please send me a separate email and I’ll point out how you can identify the problem.

Best,
Martin

3

Hi Martin,

I changed the classifier to libsvm and the decoding went fine (!) So what could be the problem when using lda?

I sent you a separate e-mail, but I’m not sure if I did it correctly. I just started using neurostars forum, so I’m sorry if I’m sending you this message twice.

Thank you!

4

Ok, it looks like there were some data points without variance, which then for the chosen Ledoit-Wolf shrinkage operation that retains the variances of the original data (shrinkage = lw2) leads to division by zero. Long story short, I have updated the version of that shrinkage operator to catch this issue. For everyone else reading this, I would try avoiding all 0 data, since other tools might not catch it.

Thanks for spotting this, Lénia!
Martin