LeanAI ASSTF

Adaptive State-Space Transfer Function

Tiny, self-adapting neural layers for efficient edge AI and continual learning.

ASSTF (Adaptive State-Space Transfer Function) is a PyTorch framework that lets neural layers continuously reconfigure their effective state space through a small set of structural parameters. It combines parameter-efficient learning with inference-time adaptation, making it ideal for edge devices, personalization, and continual learning.

---

🚀 Current Status: Stage 1 — Grounding Credit

This repository is the Stage 1 reference implementation of ASSTF. It is designed to be:

• Reproducible: every demo includes open data-generation scripts; results can be verified on a laptop.
• Extensible: drop-in ASSTFLinear / ASSTFConv1d layers let you test ASSTF on your own architectures.
• Transparent: we clearly label which demos use synthetic data and which use real-world datasets.

App 01–05 use controlled, reproducible synthetic data to isolate and demonstrate ASSTF's algorithmic behavior.
App 06 validates ASSTF on the real SST-2 sentiment dataset.

See docs/OPEN_SOURCE_STAGES.md for the full three-stage roadmap toward global adoption.

---

Why ASSTF?

• 🧠 Parameter-efficient — low-rank structural adapters add far fewer parameters than dense layers.
• 🔄 Adapts at inference time — personalize to users, noise, or drift without retraining.
• 📱 Edge-ready — runs on CPU, mobile, and embedded hardware.
• 🔒 Privacy-first — on-device adaptation; user data never leaves the device.
• 🌱 Sustainable AI — smaller models mean lower energy and memory costs.
• 🔌 Drop-in replacement — swap nn.Linear → ASSTFLinear in minutes.

---

30-Second Quick Start

import torch
from asstf import ASSTFLinear

# Replace any nn.Linear with ASSTFLinear
layer = ASSTFLinear(128, 64, structural_rank=4)
x = torch.randn(16, 128)
out = layer(x)  # shape: (16, 64)

Train and evaluate all six demos:

git clone https://github.com/SafewareTaiwan/leanai-asstf.git
cd leanai-asstf
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
pytest tests/ -v
python run_all.py

---

Installation

From PyPI (recommended)

pip install leanai-asstf

From source

git clone https://github.com/SafewareTaiwan/leanai-asstf.git
cd leanai-asstf
pip install -r requirements.txt

For NLP demos:

pip install -r requirements-optional.txt

---

How It Works

ASSTF replaces a standard layer with the composition of a core transfer function Γ and a structural modulator Ψ:

base = x · W_c^T + b_c              # core knowledge (θc)
M    = σ(gate) · (U · V)            # low-rank structural adapter (θs)
out  = base + x · M^T               # composed output

• θc stores the base task knowledge.
• θs controls how the layer adapts to new data.
• At inference time, SurpriseMinimizer updates only θs, enabling fast personalization.

Read more:

• Introduction
• Architecture & Technical Explanation
• Usage Guide
• API Reference
• Open-Source Stages

---

The 5+1 Demos

#	Application	Highlight	Data	Type
1	Embedded Gesture Recognition	Online personalization to new users	Synthetic multi-user IMU	Proof-of-concept
2	Wake-Word Detection	ASSTFConv1d + dynamic rank under noise	Synthetic 1D audio	Proof-of-concept
3	Time-Series Anomaly Detection	Continuous learning under concept drift	Synthetic industrial sensors	Proof-of-concept
4	Few-Shot Meta-Learner	θs as task-specific meta-knowledge	sklearn digits / mini-tasks	Proof-of-concept
5	Online RL Dynamic Policy	Policy adaptation to changing dynamics	Custom pendulum environment	Proof-of-concept
+1	Edge NLP	ASSTF replaces BERT-Tiny FFN layers	Real SST-2 + SentencePiece	Real-world validation

All synthetic demos include open, seeded data-generation scripts so you can reproduce them exactly and swap in your own datasets.

---

Benchmark Highlights

App	Model	Params	Metric	Data Type
Gesture	ASSTF	4,665	100.0% acc	Synthetic
Gesture	Static	38,213	100.0% acc	Synthetic
Wake-word (-5 dB)	ASSTF	779	74.62% acc	Synthetic
Wake-word (-5 dB)	Static	1,202	57.13% acc	Synthetic
Anomaly	ASSTF	506	F1 = 0.946	Synthetic
Anomaly	Static	2,080	F1 = 0.508	Synthetic
Few-shot	ASSTF	2,105	19.73% acc	sklearn digits
Few-shot	Static	7,237	19.39% acc	sklearn digits
RL adapted reward	ASSTF	965	39.66	Custom env
RL adapted reward	Static	1,441	32.97	Custom env
Edge NLP (SST-2)	ASSTF (h=96)	403,896	67.88% acc	Real
Edge NLP (SST-2)	Static	649,218	67.75% acc	Real

See docs/BENCHMARKS.md for full results and reproduction commands.

Note on interpretation: App 01–05 demonstrate ASSTF's algorithmic behavior under controlled, reproducible conditions. They are not claims of real-world SOTA. App 06 is the real-world validation. We welcome community contributions of additional real-dataset benchmarks.

---

Training a Single Demo

python app_03_anomaly/train.py
python app_03_anomaly/evaluate.py

For a quick smoke test across all apps:

python run_all.py --quick

For the full benchmark suite that matches docs/BENCHMARKS.md:

python run_all.py

---

Project Structure

leanai-asstf/
├── asstf/                  # Core ASSTF framework
├── shared/                 # Baselines, metrics, early stopping
├── app_01_gesture/         # Gesture recognition demo
├── app_02_wake_word/       # Wake-word detection demo
├── app_03_anomaly/         # Anomaly detection demo
├── app_04_few_shot/        # Few-shot learning demo
├── app_05_rl/              # Reinforcement learning demo
├── app_06_edge_nlp/        # Edge NLP demo
├── tests/                  # Unit and smoke tests
├── docs/                   # Documentation and licenses
├── data/                   # Datasets and caches
├── results/                # Benchmark outputs
├── checkpoints/            # Saved models
├── run_all.py              # Run all demos
├── requirements.txt
└── README.md

---

Roadmap & Stages

Stage	Timeline	Focus	Key Deliverables
Stage 1: Grounding Credit	Now – Month 4	Reproducible reference implementation + community	This repo, PyPI, CI, launch
Stage 2: Real-World Value	Month 4–12	Public benchmarks, integrations, academic paper	Real datasets, Hugging Face, Edge Impulse, arXiv
Stage 3: Commercial Value	Month 12–24	Enterprise tools, platform, support	LeanAI Studio, Edge SDK, managed inference

See docs/OPEN_SOURCE_STAGES.md for the complete plan.

---

License

This project is dual-licensed:

• Community License — free for personal learning, academic research, non-profit education, and open-source contributions.
• Commercial License — required for any commercial product, service, SaaS, cloud deployment, or hardware integration.

See LICENSE and docs/COMMERCIAL_LICENSE.md for details.

For licensing inquiries, contact Bentley@safeware.com.tw.

---

Citation

If you use ASSTF in academic research, please cite the white-paper:

@techreport{lin2025asstf,
  title={Mathematical Framework to Enable Adaptive Neuron Transitions},
  author={Lin, Bentley Yusen},
  institution={Safeware Technologies Inc., Ltd.},
  year={2025}
}

---

Contributing

We welcome research contributions, bug reports, documentation improvements, and especially real-dataset benchmarks. Please see CONTRIBUTING.md for guidelines.

By contributing, you agree that your contributions may be used under both the Community License and future commercial licensing.

---

Acknowledgments

ASSTF is developed by Safeware Technologies Inc., Ltd. The reference implementation is intended for research and demonstration; production deployment requires additional tuning and a commercial license.

---

Keywords: Adaptive State-Space Transfer Function, ASSTF, LeanAI, parameter-efficient deep learning, edge AI, TinyML, continual learning, test-time adaptation, online personalization, low-rank adaptation, PyTorch, efficient AI, green AI, on-device ML.