Skip to content

How stable are untrained CNNs? Measuring random-weight CNN variability on EEG with Fuzzy PyTorch #4

Open

Description

@royjoysoy

Title

How stable are untrained CNNs? Measuring random-weight CNN variability on EEG with Fuzzy PyTorch

Leaders

Roy Seo (post-doc, Centre for Addiction and Mental Health / Krembil Centre NeuroInformatics)
Email: seoroy15@gmail.com
Discord: royjoysoy

Collaborators

Yohan Chatelain (scientific associate, Centre for Addiction and Mental Health / Krembil Centre NeuroInformatics)
Email: Yohan.Chatelain@camh.ca

Mina Alizadeh (Ph.D. student)
Email: mina.alizadeh94@gmail.com

Injy Fouda (Master's student)
Email: injyfouda1@yahoo.com

Project Description

Untrained CNNs use random weights. They are never trained. Yet they can still extract useful features from data, including EEG.

We want to understand them better. Our main question is simple: how much do untrained CNNs vary?

They can vary in two ways:

  • Different random weights (different seeds).
  • Small rounding errors in computer arithmetic.

To measure the second kind, we use Fuzzy PyTorch. It adds tiny random noise to floating-point math. This is called Monte Carlo Arithmetic (MCA). We run the same model many times and compare the outputs.

We will use open EEG datasets.

Who is this for?

  • Anyone curious about random weights and untrained CNNs.
  • Anyone who wants to know why they work on EEG, or why they don't.
  • Anyone who knows good open EEG datasets. Please tell us!

I am new to untrained CNNs too. We will learn together. By the end of Brainhack, we will all know a lot more.

Our goal is to write a short paper together. Everyone who helps will be credited.

Key resources:

  • What is Fuzzy? : https://github.com/verificarlo/fuzzy
  • Fuzzy PyTorch: Gonzalez-Pepe, Akhaddar, Glatard & Chatelain (2026). "Fuzzy PyTorch: Rapid Numerical Variability Evaluation for Deep Learning Models." TMLR. https://openreview.net/forum?id=0ogq232VGP
  • Verificarlo (the tool behind Fuzzy PyTorch): Denis, de Oliveira Castro & Petit (2016). "Verificarlo: Checking Floating Point Accuracy through Monte Carlo Arithmetic." IEEE ARITH.

Link to project repository/sources

TBD

Goals for Brainhack Montreal

  • Beginner: Find and list open EEG datasets we can use.
  • Beginner: Run an untrained CNN on one EEG dataset.
  • Intermediate: Compare results across different random seeds.
  • Intermediate: Run the same model with Fuzzy PyTorch (MCA) and measure the variability.
  • Advanced: Compare seed variability with MCA variability.
  • Final: Write a short paper draft together.

Skills

  • Python: run and change simple scripts (beginner)
  • EEG: find and load open datasets, e.g., with MNE (beginner)
  • PyTorch: build and run small CNNs (intermediate)
  • Statistics: measure and compare variability (intermediate)
  • Writing: help write and edit the paper (beginner, non-coding)

Onboarding documentation / Expectations

TBD. Join for a few hours or all 3 days. Any time helps.

What will participants learn?

  • What untrained (random-weight) CNNs are and why they matter.
  • What numerical variability is and how to measure it.
  • How to use Fuzzy PyTorch.
  • How to work with open EEG data.

Data to use

Open EEG datasets. Still TBD. Do you know a good one? Please tell us or join the team!

Credit to collaborators

All contributors will be listed in the README. People who contribute to the paper will be invited as co-authors.

Image

No response

Type

methods development

Project Maturity Status

1 - prototype

Topic

machine_learning

Tools

Machine_learning

Programming language

Python

Modalities

EEG

Git skills

1 - commit and push

Anything else?

Nice to meet you : )

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions