Saved in:
Bibliographic Details
Main Author: Joneidi, Mohsen
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2510.09825
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914087522271232
author Joneidi, Mohsen
author_facet Joneidi, Mohsen
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2510_09825
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Decomposer Networks: Deep Component Analysis and Synthesis
Joneidi, Mohsen
Machine Learning
Computer Vision and Pattern Recognition
Information Theory
Neural and Evolutionary Computing
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.
title Decomposer Networks: Deep Component Analysis and Synthesis
topic Machine Learning
Computer Vision and Pattern Recognition
Information Theory
Neural and Evolutionary Computing
url https://arxiv.org/abs/2510.09825