Ko fi https://ko-fi.com/utah23

Please ask for paypal

IF YOU WANT TO FUND ME TO BE INCLUDED ON PAPERS THAT HAVE NOT BEEN PUBLISHED YET, PLEASE MESSAGE ME OR EMAIL ME AT utah@utahcreates.com AND I WILL PUT YOU DOWN ON THE FUNDING SECTION. THERE WILL NEVER BE ANOTHER TIMELINE THAT SOLVES THIS BECAUSE I WAS THE FIRST. BE PART OF HISTORY BEFORE THEY GET PUBLISHED.

https://doi.org/10.5281/zenodo.18107992

Liquid-Potato-Edge

Biomimetic Liquid Neural Architectures Enable Robust Control on Legacy Edge Silicon without Accelerators

A compact research code and manuscript demonstrating that small, continuous-time "Liquid" neural networks can deliver robust control and lower-latency inference on commodity CPU-only edge hardware (e.g., Intel Celeron N5095), with no GPU/TPU/NPU accelerators.

---

Table of Contents

• Overview
• Key Results
• Repository Structure
• Quick Start
  • Run the demo model
  • Compile the manuscript (PDF)
  • Package for arXiv
• Data and Figures
• Reproducibility Notes
• Citing
• License
• Contact

---

Overview

Liquid Time-Constant (LTC) networks, formulated as continuous-time dynamical systems, can outperform discretized recurrent baselines on CPU-only edge hardware. We deploy a sparse, 14‑neuron Liquid model on a 15 W Intel Celeron N5095 edge node using strictly CPU instruction sets and observe:

• a 5.8% reduction in inference latency, and
• a 25.8% increase in robustness to Gaussian input noise

relative to an LSTM baseline. These findings indicate that differential neural architectures provide a viable path to robust, low-latency autonomous control on localized, low-power silicon.

Figure 1. Liquid architecture shows lower latency and higher stability under Gaussian noise compared to an LSTM baseline.

Key Results

• Hardware: Intel Celeron N5095 (Jasper Lake), CPU-only, ~15 W TDP
• Latency (avg inference): Liquid 0.317 ms vs. LSTM 0.337 ms → 5.8% faster
• Parameters: ~400 (Liquid) vs. ~17,000 (LSTM) → ~97% smaller
• Robustness to noise: 24.86 steps (Liquid) vs. 19.76 steps (LSTM) before failure (+25.8%)
• Software: Python + PyTorch (CPU); manuscript in LaTeX with BibTeX

Repository Structure

• manuscript.tex — Main LaTeX paper (modernized preamble, numbered sections, robust cross-referencing)
• sample.bib — Minimal BibTeX with a placeholder for the LTC citation (hasani2021)
• FIGURE_1.png — Figure used in the paper (case-sensitive on Linux/arXiv)
• paper_data.csv — Supplemental data used to create figures or tables (optional)
• POTATO_AI.py — Minimal PyTorch implementation of a 14-neuron Liquid cell demo
• ARXIV_LAUNCHER.py — Helper script to package sources and open the arXiv submission portal
• arXiv_Submission_Package.tar.gz — Output produced by the launcher after packaging (generated)

Quick Start

Prerequisites

• Python 3.9+ (3.10/3.11 also fine)
• pip (or conda)
• PyTorch (CPU build is sufficient)

Install PyTorch (CPU) with pip:

# Generic CPU-only install (Linux/Mac)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu

# Windows alternative (if the above fails)
pip install torch torchvision torchaudio

Note: You do NOT need CUDA/ROCm/GPU drivers. All results are CPU-only.

Run the demo model

python POTATO_AI.py

Expected console output includes model activation lines like:

// POTATO BRAIN ONLINE.
// NEURON COUNT: 14
// ENERGY CONSUMPTION: NEGLIGIBLE.
PotatoBrain(
  (liquid_layer): LiquidCell(...)
  (readout): Linear(in_features=14, out_features=1, bias=true)
)

Compile the manuscript (PDF)

Use standard LaTeX/BibTeX:

pdflatex manuscript.tex
bibtex manuscript
pdflatex manuscript.tex
pdflatex manuscript.tex

Or with latexmk:

latexmk -pdf manuscript.tex

Files required for a clean compile:

• manuscript.tex
• sample.bib
• FIGURE_1.png

Package for arXiv

The arXiv prefers source files; this script bundles what you need and opens the submission page.

python ARXIV_LAUNCHER.py

It will:

1. Verify required files exist
2. Create arXiv_Submission_Package.tar.gz
3. Open your browser at the arXiv submission portal

Tip: Filename case matters on Linux/arXiv. Ensure the figure is named exactly FIGURE_1.png.

Data and Figures

• paper_data.csv provides transparency for figure/table values. You can explore it with pandas:

import pandas as pd
import matplotlib.pyplot as plt

df = pd.read_csv('paper_data.csv')
print(df.head())
# Example: if latency columns exist, plot them
cols = [c for c in df.columns if 'latency' in c.lower() or 'steps' in c.lower()]
df[cols].plot(kind='bar')
plt.title('Paper metrics')
plt.tight_layout()
plt.show()

Reproducibility Notes

• All experiments were run on CPU-only (Intel N5095). CUDA/ROCm were disabled.
• The Liquid cell in POTATO_AI.py is a minimal, readable demonstration of the continuous-time dynamics used in the manuscript; it is not a full training/reproduction pipeline. If you want a training/evaluation script added, please open an issue.
• Randomness: for rigorous experiments, set seeds and control non-determinism where applicable.

Citing

If this repository or manuscript is useful in your research, please cite:

@misc{hans2025_liquid_potato_edge,
  title        = {Biomimetic Liquid Neural Architectures Enable Robust Control on Legacy Edge Silicon without Accelerators},
  author       = {Hans, Utah},
  year         = {2025},
  howpublished = {GitHub repository},
  url          = {https://github.com/utahisnotastate/Liquid-Potato-Edge}
}

If you reference Liquid Time-Constant networks, please also cite:

@article{hasani2021,
  title        = {Liquid Time-Constant Networks},
  author       = {Hasani, Ramin and Lechner, Mathias and Amini, Alexander and Rus, Daniela},
  journal      = {arXiv preprint arXiv:2006.04439},
  year         = {2021},
  url          = {https://arxiv.org/abs/2006.04439}
}

License

No explicit license is provided yet. All rights reserved by the author unless a license file is added. If you plan to use this work beyond fair academic use, please contact the author.

Contact

• Author: Utah Hans
• Email: utah@utahcreates.com
• Project URL: https://github.com/utahisnotastate/Liquid-Potato-Edge

---

Made with curiosity and a small, liquid brain.