PAULA: Predictive Adaptive Unsupervised Learning Agent

Empirical Validation of the ALERM Framework

Part of the Artificial Life Research Initiative.

📄 Research Paper: ALERM Framework Draft (Coming Soon)
📖 Read the article: Brain-Inspired AI: Early Results from a Radical New Neuron Model

This repository contains the implementation of PAULA (Predictive Adaptive Unsupervised Learning Agent). It is a radically different approach to neural network modeling designed explicitly to serve as the concrete empirical validation for the ALERM (Architecture, Learning, Energy, Recall, Memory) mathematical framework.

Unlike traditional deep learning systems that rely on backpropagation, global objective functions, and dense weight matrices, PAULA achieves 84.3% accuracy on MNIST using:

• Purely local learning rules (no backpropagation).
• Extreme sparsity (25% connectivity, 5-10% active neurons per layer).
• Fast homeostatic metaplasticity (tick-by-tick adaptive thresholds).
• Temporal/Spiking dynamics (Time-to-first-spike inference mechanisms).

This isn't just another neural network implementation—it's a virtual laboratory where we can experiment with the very building blocks of the mind.

---

🔬 Reproducing the ALERM Validation (For Researchers)

If you are coming from the ALERM/PAULA papers, this repository contains the exact codebase used to generate the empirical ablation studies. We provide a direct mapping from the paper's theoretical concepts to the physical code components.

Code Alignment with ALERM Framework

• Architecture (A) ↔ Memory (M): Sparsity isn't just for efficiency; it is necessary for dynamical stability in PAULA. Dense networks fail to reach homeostatic equilibrium.
  • Code location: network.py, networks/ (where the 6-layer, 144-neuron sparse architecture is generated).
• Learning & Plasticity (L): Uses a triad of rules: Hebbian Association, Homeostatic Stability, and Predictive Error Minimization.
  • Code location: neuron.py (core update logic), neuron/ablation_registry.py (isolated mechanism testing).
• Energy (E) ↔ Recall (R): Temporal inference follows a Drift-Diffusion model framework, where fast decisions (top-3 candidates) cost minimal energy, but complex patterns cause the system to deliberate, costing more spikes (energy) for higher accuracy.
  • Code location: evals/, experiments/

Running the Core Experiments

To reproduce the core 84.3% MNIST evaluation or run the ablation models described in the paper, please refer strictly to our dedicated Reproducibility Guide (Contains precise instructions to replicate the findings).

---

🛠 The Virtual Laboratory (For Contributors)

While the core scripts validating the ALERM paper are the primary focus of this repository, we also ship tools for interactive, hands-on experimentation.

Quick Start / Installation

The system is designed for immediate experimentation. We strongly recommend using uv for incredibly fast dependency management:

# Clone the repository
git clone https://github.com/arteriali/neuron-model.git
cd neuron-model

# Install the project (makes all scripts runnable from project root)
uv pip install -e .

# Or install dependencies only
uv pip install -r requirements.txt

Legacy & Experimental Components (⚠️ Warning)

This repository evolved rapidly from a one-man obsession into a formal research platform. As a result, there are several tools in the repository that are highly useful but are currently considered unstable/experimental:

• CLI (cli/) & Web Viz (web_viz/): We built a rich, interactive command-line interface and real-time visualization layer (WebSocket + Cytoscape.js). They are functional and highly useful for exploring network topologies and attractor dynamics. However, they are unstable, experimental, and may contain bugs since they are not the primary focus of recent active development.
• Pipeline (pipeline/): This is an asynchronous job runner designed to batch experiments. It is half-functional but highly useful conceptually. Contributions to finish and stabilize this component are actively welcomed!
• Modality Processor (modality_processor/): This module is an untested stub meant for multi-modal sensory input (video, audio). It is not part of the core validated framework.

---

📊 Key Discoveries & Ablation Results

The PAULA implementation empirically proves several novel predictions generated by the ALERM framework.

Summary of our ablation studies:

Model Configuration	Phase-Space Dynamics	Performance (MNIST)	Implication
Baseline (25% Sparse + Adaptive t_ref)	Stable Limit Cycle	84.3%	Optimal, energy-efficient stability.
Dense (100% Connectivity)	Chaotic Fluctuations	N/A (Failed)	Sparsity is required for stable memory.
Frozen Modifiers (No Homeostasis)	Catastrophic Variance	80.4%	Homeostasis provides fast metaplasticity.

Complex datasets (like distinguishing a messy '8' from a '9') require multi-scale temporal logic and deliberate time-to-first-spike evaluation; simple patterns naturally execute fast hypotheses.

---

🤝 Join the Journey

Here's the truth: this started as a one-man obsession, but it's grown into something bigger than what I can do alone. I've spent countless hours developing the theoretical foundations and turning abstract ideas into working models, but the really exciting discoveries—the ones that will change how we think about intelligence—those will come from all of us working together.

If you're a researcher: Run experiments I haven't thought of. Push the model to its limits. Reference the REPRODUCIBILITY.md guide to start building ablations. If you're a developer: Help me make this tool more powerful. Stabilize the pipeline module, fix up the CLI and web visualization bugs, or optimize the core loop.

Found something broken? Open an issue. Want to collaborate? Reach out!

• See the bigger picture: al.arteriali.st
• Direct collaboration: Reach out via my GitHub profile contacts.