Disentangling signal and noise in neural responses through generative modeling
Fig 7
GSN disentangles signal and noise in principal components analysis (PCA).
Here we use PCA to analyze the results of GSN as applied to FFA-1 (code available at https://osf.io/f34bc). A, Eigenspectra. For each of the eight participants (P1–P8), we plot the eigenspectra of the signal and noise as estimated by GSN (‘Signal (GSN)’, ‘Noise (GSN)’), as well as the eigenspectrum of the trial-averaged data (‘Naive’). The main plots show results on a linear scale for up to the first 10 dimensions; the insets show results on a base-10 log-log scale for up to the first 100 dimensions. Numbers above each main plot indicate the effective dimensionality of the three eigenspectra. B, Split-half reliability of principal components. The cosine similarity between corresponding principal components from two split-halves of the data from each participant is plotted for up to the first 100 dimensions. The thick black line indicates the mean across participants. C, Across-participant consistency. A common set of 515 images were viewed three times each by all participants. For each participant, we computed the projections of trial-averaged responses to these 515 images onto either (i) the first principal component of the covariance of the trial-averaged data (‘Standard PCA’) or (ii) the first principal component of the signal covariance estimated by GSN (‘GSN PCA’). The cosine similarity of these projections between each pair of participants is shown.