High-speed feature extraction with integrated microwave neurons

This repository contains the full experimental data, simulation code, and analysis pipeline.

Reproducibility Notes

• Please select thresholds used for compatibility and association matrices' alignment balance sparsity and overlap; moderate variations do not change qualitative outcomes
• Windows paths are currently hard-coded in several scripts and must be updated locally
• GPU acceleration is optional but recommended for training stages
• The inference was conducted by feeding the mapped pulse-train tokens to the MNN, interfacing the .pth Python file trained linear with MATLAB's data acquision. However, in absence of the phsycial chip, readers can use the simulated pipelines by using cached (pre-recorded) MNN waveforms. 
• Extract all the dataset(s).

Subroutines:

1. Run submarine game data generation

2. Process MNN waveform measurements

3. Train association matrix

4. Perform state–token alignment

5. Train and test submarine game models

6. Run baseline comparisons

7. Execute static-ratio image reconstruction

8. Run sentence-building pipeline