Project Overview

Perforated DreaMS, an optimization of the Deep Representations of Mass Spectra (DreaMS) foundation model. Successfully compressed a scientific Transformer model to run on low-power edge hardware (like Raspberry Pi) for real-time chemical identification in clinical labs.

Key Results (The "Rule of Three")

Compared the Giant Baseline against a Standard Compressed model and our Perforated Model on a blind test of industrial chemical spectra.

Technical Highlights

• ** 3D Patch:** implemented a custom fix to apply Perforated Backpropagation to 3D Transformer tensors, enabling dendritic optimization on sequence data.
• Industrial Augmentation: Trained on GNPS Metabolomics data with 20x stochastic augmentation to simulate real-world signal noise.
• Edge Ready: The model now fits in the L2 Cache of standard ARM processors, eliminating the need for cloud computing in mass spectrometry.

Submission Checklist

• README.md (Follows official template)
• main.py (Reproducible code with library patches included)
• preforated-dreams.ipynb (Full research notebook)
• requirements.txt
• Mandatory Graph: PAI.png (Shows the Dendritic Spike)
• Supplementary PAI Graphs: PAI_1.png, PAI_2.png
• Bonus Graphs: Robustness_Curve.png, Latency_Benchmark.png, Dendritic_Latent_Space.png
• Summary Visual: DreaMS_Submission_Graph.webp

Thanks for review!