Invariance analysis

To determine whether the pattern of neural activity to different intonation contours is the same across different data sets (e.g. speech data vs. nonspeech control data), we used linear discriminant analysis (LDA) to fit models to predict intonation contour (i.e. the intonation/sentence_type condition between 1-4) from the neural activity time series from each electrode. These models were fit on one data set and tested on another, and the performance accuracies of the model on different test sets were compared.

intonatang.intonation_invariance.load_control_test_accs(subject_number, chans=None, diagonal=True, missing_f0=False, zscore_to_silence=True)

Used to load nonspeech control invariance analysis results.

Parameters:
  • subject_number – xxx in ECxxx
  • diagonal (bool) – whether the invariance analysis used a shrinkage of 1 and was diagonal LDA.
Returns:

  • accs (ndarray): shape is (n_chans x n_perms x 3). The last dimension contains accuracy values for held out

    speech data, shuffled speech data, and shuffled nonspeech data

  • accs_test (ndarray): shape is (n_chans). Contains accuracy value for nonspeech data.
Return type:

(tuple)
intonatang.intonation_invariance.save_control_test_accs(subject_number, accs, accs_test, chans=None, diagonal=True, missing_f0=False, zscore_to_silence=True)

Used to save the nonspeech control invariance analysis results.

After running test_invariance_control, save accs and accs_test for a subject.

Parameters:
  • subject_number – xxx in ECxxx
  • accs (ndarray) – shape is (n_chans x n_perms x 3). The last dimension contains accuracy values for held out speech data, shuffled speech data, and shuffled nonspeech data
  • accs_test (ndarray) – shape is (n_chans). Contains accuracy value for nonspeech data.
  • diagonal (bool) – whether the invariance analysis used a shrinkage of 1 and was diagonal LDA.
  • missing_f0 (bool) – missing f0 invariance analysis
intonatang.intonation_invariance.test_invariance_control(subject_number, solver='lsqr', shrinkage=1, n_perms=1000)

Run the LDA invariance analysis on nonspeech control data.

This function contains the pipeline for the nonspeech invariance analysis. Here, we use LDA to fit a model on the neural activity time series to predict intonation condition using speech data. We then test this model on both held out speech data and the nonspeech control data. The accuracies for the model on the held out speech data and nonspeech data are then returned and can be saved with save_control_test_accs.

Parameters:
  • subject_number – xxx in ECxxx
  • solver – The type of solver to use for LDA. Can be “svd” or “lsqr”. Only “lsqr” supports shrinkage/regularization.
  • shrinkage – The shrinkage parameter between 0 and 1. Default of 1 is diagonal LDA.
  • n_perms – The number of permutations for held out speech data to run.
Returns:

  • accs (ndarray): shape is (n_chans x n_perms x 3). The last dimension contains accuracy values for held out

    speech data, shuffled speech data, and shuffled nonspeech data

  • accs_test (ndarray): shape is (n_chans). Contains accuracy value for nonspeech data.
Return type:

(tuple)