HSAMA: Hierarchical Symbolic Meta-Architecture

This repository contains the core implementation of the HSAMA architecture (Hierarchical Symbolic Meta-Architecture) adapted to operate with Multi-Layer Perceptron (MLP) interconnections with dynamic biases generated by hypernetworks (MLP Edge with Dynamic Bias).

---

Repository Structure

The architecture is organized into three main layers:

• Models (src/models/): Definement of the neural graph models (HSAMA topology, KAN, MLP Edge with Dynamic Bias).
  • graph.py: Defines the neural graph structures and K-regular graph topologies.
  • hsama.py: Core mathematical definition and forward propagation of the HSAMA model.
  • kan.py: Pade KAN layer and network implementations.
• Runtime & Memory (src/runtime/): Online runtime execution, temporal memory buffers, and surprise estimators.
  • multiscale.py: Multi-scale temporal memory components (T0/T1 scale encoders).
  • online.py: Online runtime loop coordinating graph updates and inference.
  • replay.py: Surprisal-based prioritized experience replay buffer.
  • surprisal.py: Surprise estimators adjusting model dynamics on regime changes.
  • temporal.py: Temporal history ring buffers.
• Learning & Optimization (src/learning/): Optimization objectives.
  • objectives.py: Causal loss formulations, Huber objectives, and Sharpe/Drawdown proxies.

---

Dependencies

• Python >= 3.12
• PyTorch >= 2.2
• NetworkX >= 3, < 4