We introduce TSNet (Topological Structure Networks), a gradient-free visual architecture grounded in the hypothesis that images encode structural evidence of their underlying generative process. The model extracts multi-scale path-based and deterministic topological features, and performs classification via a relational density field without gradient-based optimization, while producing an intrinsically grounded epistemic confidence signal. On MNIST, TSNet achieves competitive accuracy (0.832 standalone; 0.876 in hybrid mode) while demonstrating strong epistemic behavior, including 99.5% rejection of out-of-distribution samples on Fashion-MNIST compared to 37% for an MLP. It also shows high efficiency in low-data regimes (CDR@5 = 0.968), indicating that structural representations are compact and information-dense. TSNet constitutes the first visual-domain instantiation of the AI Implicit paradigm, where epistemic awareness emerges directly from structural representation rather than post-hoc calibration.
Authors: Momen Ghazouani
DOI: https://doi.org/10.5281/zenodo.19928909
Publish Year: 2026
Download PDF