Learn Z0 ≡ color and Z1 ≡ size. Latent dimension Z0 is passed to the classifier for color and latent dimension Z1 is passed to the classifier for size. The user-uttered labels for color are red, blue and for size are small, big. The color labels are assigned to a single variation of the respective color. All images which can not be described by the given labels—yellow and green for color and medium for size—are given an unknown ground truth label. Total |Z| = 4.
Scatter plot of the validation data points for concept group 0 (color) over Z0 (x axis) and Z1 (y axis)
Scatter plot of the validation data points for concept group 1 (size) over Z0 (x axis) and Z1 (y axis)
Per-concept group confusion matrix for classifying every image part of the customised dSprites-based dataset. The labels are consistent with the the description at the top of the page. The statistics from the matrices are used to calculate the reported F1 scores in the paper.
Cosine similarity diagrams for the labels from concept group 0 (blue, red) and concept group 1 (big, small). The average entropy of the distributions of normalised cosine values for the white cells, for each label, is reported in the paper. The value below each component from the PCA analysis represents the eigen value for that component.
Scatter plot of the validation data points for concept group 0 (color) over Z0 (x axis) and Z1 (y axis)
Scatter plot of the validation data points for concept group 1 (size) over Z0 (x axis) and Z1 (y axis)
Per-concept group confusion matrix for classifying every image part of the customised dSprites-based dataset. The labels are consistent with the the description at the top of the page. The statistics from the matrices are used to calculate the reported F1 scores in the paper.
Cosine similarity diagrams for the labels from concept group 0 (blue, red) and concept group 1 (big, small). The average entropy of the distributions of normalised cosine values for the white cells, for each label, is reported in the paper. The value below each component from the PCA analysis represents the eigen value for that component.
Scatter plot of the validation data points for concept group 0 (color) over Z0 (x axis) and Z1 (y axis)
Scatter plot of the validation data points for concept group 1 (size) over Z0 (x axis) and Z1 (y axis).
Per-concept group confusion matrix for classifying every image part of the customised dSprites-based dataset. The labels are consistent with the the description at the top of the page. The statistics from the matrices are used to calculate the reported F1 scores in the paper.
Cosine similarity diagrams for the labels from concept group 0 (blue, red) and concept group 1 (big, small). The average entropy of the distributions of normalised cosine values for the white cells, for each label, is reported in the paper. The value below each component from the PCA analysis represents the eigen value for that component.