Zero-Retraining Topological Targeting for Edge AI: Exact Class-Selective Specialization via König Bipartite Matching
Authors/Creators
Description
We present an exact, sub-millisecond algorithm for task-specific neural network specialization requiring no retraining, no gradients, and no active neural network to execute. By modeling the Fully Connected layer as a bipartite graph, we apply König's Theorem to compute the Minimum Vertex Cover in polynomial time via Hopcroft-Karp matching — identifying the exact minimal set of neurons required for any target vocabulary. Unlike all existing pruning methods, which require the neural network to prune itself via backpropagation, this approach operates exclusively on the static weight matrix using pure combinatorial mathematics. Combined with dynamic per-class thresholding and Shannon-stratified sampling, the algorithm specializes a production MobileNetV2 classifier to any target class in a single mathematical pass, producing a 2,280-byte deployment artifact in sub-second time on a commercial Snapdragon 8 Gen 2 processor — without GPU, server, or framework dependency. We validate targeted recovery of suppressed classes, including a Siamese Cat naturally absent from blind pruning, achieving Top-1 classification after 84.6% structural compression (p = 0.000204). The method is hardware-agnostic and maps directly onto FPGA and programmable edge silicon. This work establishes prior art for zero-retraining, gradient-free topological specialization of foundation models for edge deployment.
Notes
License Amendment Note: > Please note that the "All rights reserved" footer appearing in the PDF manuscript is a legacy placeholder. This work is formally distributed under a dual-licensing scheme: the manuscript text is licensed under CC BY-NC-ND 4.0, and the associated algorithmic methods and source code are licensed under the PolyForm Noncommercial License 1.0.0. The terms of these licenses supersede any "All rights reserved" notice within the document.
LEGAL NOTICE & INTELLECTUAL PROPERTY WARNING
License: PolyForm Noncommercial License 1.0.0
Author & Rightsholder: Andrés Sebastián Pirolo
The exact topological pruning algorithm, mathematical frameworks, and underlying concepts detailed in this publication and accompanying artifacts are strictly licensed under the PolyForm Noncommercial License 1.0.0.
1. Hardware and Software Implementation Restrictions:
Any unauthorized translation, adaptation, or implementation of this algorithm into any programmable hardware (including but not limited to FPGAs, ASICs, SoCs, and Edge silicon) or software architecture for commercial, proprietary, or industrial purposes is strictly prohibited.
2. International Enforcement and Litigation:
This intellectual property is protected under international copyright treaties and intellectual property conventions. The author explicitly reserves the exclusive right to initiate international litigation, seek immediate injunctions, and pursue financial damages against any individual, corporation, or entity that executes unauthorized commercial implementations without prior written consent.
The disclosure of these concepts serves as Prior Art to prevent third-party patenting, but does not grant commercial usage rights.
For commercial licensing and authorized implementation inquiries, contact: andrespirolo@gmail.com
Files
Pirolo2026Konig_V2.pdf
Files
(2.0 MB)
| Name | Size | Download all |
|---|---|---|
|
md5:c39779a626a7a4e1e1019a19440ba312
|
1.0 MB | Preview Download |
|
md5:54a691ad74ada25385342ac507ed9179
|
1.0 MB | Preview Download |
Additional details
Software
- Programming language
- C++