Fig. 2
Assessing algorithm and dimensionality stability with singular vector canonical correlation analysis (SVCCA). a SVCCA applied to the weight matrices learned by each compression algorithm in gene expression data from The Cancer Genome Atlas (TCGA). The mean of all canonical correlations comparing independent iterations is shown. The distribution of mean similarity represents a comparison of all pairwise iterations within and across algorithms. The upper triangle represents SVCCA applied to real gene expression data, while the lower triangle represents permuted expression data. Both real and permuted data are plotted along the diagonal. b Mean correlations of all iterations within algorithms but across k dimensionalities. SVCCA will identify min(i, j) canonical vectors for latent dimensionalities ki and kj. The mean of all pairwise correlations is shown for all combinations of k dimensionalities