Decomposer Networks: Deep Component Analysis and Synthesis
This provides a novel approach for interpretable representation learning in machine learning, though it appears incremental relative to existing object-centric architectures.
The authors tackled the problem of creating interpretable component representations by proposing Decomposer Networks (DecompNet), a semantic autoencoder that factorizes inputs into multiple components using a residual update rule, achieving parsimonious and semantically meaningful representations.
We propose the Decomposer Networks (DecompNet), a semantic autoencoder that factorizes an input into multiple interpretable components. Unlike classical autoencoders that compress an input into a single latent representation, the Decomposer Network maintains N parallel branches, each assigned a residual input defined as the original signal minus the reconstructions of all other branches. By unrolling a Gauss--Seidel style block-coordinate descent into a differentiable network, DecompNet enforce explicit competition among components, yielding parsimonious, semantically meaningful representations. We situate our model relative to linear decomposition methods (PCA, NMF), deep unrolled optimization, and object-centric architectures (MONet, IODINE, Slot Attention), and highlight its novelty as the first semantic autoencoder to implement an all-but-one residual update rule.