BSc Thesis — Spectral Compression for Scalable Weight Space Learning

Weight space metanetworks normally read a trained network's raw weights. In this project we feed them a smaller, SVD-compressed form of those weights instead.

Full version (NKUA Pergamos): https://pergamos.lib.uoa.gr/item/uoadl:5421096

Main components

• thesis.tex, build/thesis.pdf — the thesis
• weight_space_spectral_compression.ipynb — all the experiments
• results/ — the result tables reported in the thesis

Build the thesis

latexmk -xelatex thesis.tex

Needs a XeLaTeX toolchain (TeX Live) with the CMU Bright fonts from fonts/.

Reproduce the experiments

Open the notebook and run it top to bottom. It was developed on Kaggle, so the data paths and checkpoints assume that environment — adjust the input/output directories to run it locally.