Figure 16
From: Learning without loss
Comparison of a conventional autoencoder (top row) and the proposed model based on iDE codes (bottom row), when the data is not a connected set—rendered as the Xor of two disks. In the conventional design, decoding will have interpolation artifacts (purple region of “Venn diagram”). This is avoided by iDE encoding, which only seeks to envelop the data (green boundary curve) while also constraining the representation to be disentangled (tensor product rendered as a green square). The generative model for iDE codes requires, in addition, a classifier \(\mathcal{C}\) trained to identify codes that correspond to data. When the data codes occupy half of the code space, as in this cartoon, the best setting of the fpa training parameter would be \(1/2\)