Figure 3. Improved interpretability in deep networks.

The trained reconstruction algorithm can be mapped back into its original interpretation. Hence, we can compare them to reconstruction weights after (a) Parker [17] and (b) Schäfer [18]. (c) expresses significant similarity to (b) which is also able to compensate for the loss of mass. While (b) was only arrived at heuristically (c) can be shown to be data optimal here.