HSPMN vs Transformer

Purpose

This repository provides a comparison between:

• HSPMN (Hierarchical Shallow Predictive Matter Networks)
• Transformer (standard multi-head self-attention)

For the main HSPMN repository, see: https://github.com/NetBr3ak/HSPMN

Files

1. compare.py - trains both models on identical data and reports MSE, FLOPs, parameters, and latency.
2. verify.py - verification suite to check fairness, FLOPs counting, parameter counts, gradient flow, and reproducibility.

Quick Start

If using a POSIX shell (bash/zsh):

source venv/bin/activate
pip install -r requirements.txt
python compare.py
python verify.py

If using Windows PowerShell:

./venv/Scripts/Activate.ps1
pip install -r requirements.txt
python .\compare.py
python .\verify.py

Scripts emit summary plots (loss_curves.png, flops_comparison.png, mse_comparison.png, latency_comparison.png) and the latest metrics JSON under artifacts/.

Visual summary (latest run)

Fairness and Metrics

• Identical data, optimizer (Adam, lr=0.001), epochs (100), batch size (128), loss (MSE), seed (42)
• Metrics: MSE (accuracy), FLOPs (compute), Parameters (size), Latency (inference time)

Task

Multi-scale time series forecasting (sum of three sinusoids + noise). Predict at horizons t+1, t+5, t+20 using input length 64.

Representative Results (seed=42)

• HSPMN: MSE=0.00322, Params=0.41M, FLOPs=0.41M, Latency≈0.58ms
• Transformer: MSE=0.00423, Params=0.62M, FLOPs=40.92M, Latency≈0.23ms

Interpretation: HSPMN achieves lower MSE and ~101× fewer FLOPs on this task, but is slower in wall-clock time due to limited parallelism in the current implementation. Results are reproducible and verified by verify.py.

Notes on Implementation

• HSPMN implements: hierarchical predictive modules, parallel shallow processing, non-reciprocal routing, and oscillatory synchronization (additive modulation)
• Dynamic connectivity weights are learned via backprop in this implementation; explicit active-matter update rules are planned as future work

License

MIT License.