Core Basics in ML, Signal Processing, and More August 2025 | EURECOM, Nice, France

Inspired by The Fall by Albert Camus, this project looks at how vectors, the basic building blocks of machine learning, form patterns, signals, and representations, and how these can help us understand structure in data.

Short Tutorial on Cross-Spectrum and Coherence in EEG 2018 | WAVES Group, Ghent University, Ghent, Belgium

I first discussed cross-spectrum and coherence in EEG in my 2018 PhD progress report. This document provides a short introduction to the topic.

EEG Time-Frequency Representation for Envelope Following Responses (EFR) 2018 | WAVES Group, Ghent University, Belgium

This tutorial is part of the supplementary material for my PhD thesis. It covers EEG time-frequency representations in the context of the Envelope Following Response (EFR), also known as the Frequency Following Response (FFR). Topics include induced and evoked responses, as well as phase-locked and non-phase-locked power.

Git and GitHub for Research Labs September 2026 | Columbia University & NYSPI, NYC, USA

A practical tutorial on using Git and GitHub in research projects. It covers the basic Git workflow, SSH setup, version control, and collaboration. It also includes MATLAB scripts for automating some of the daily Git operations.

The repository contains the tutorial, MATLAB code, and a PDF version.

Direction Invariant Inference for Heterogeneous Power Effects
October 2026 | Columbia University & NYSPI, NYC, USA

A statistical approach for studying condition effects in EEG when their direction varies across participants. For example, power or frequency may increase in some participants and decrease in others. When these differences are averaged, positive and negative changes can cancel each other out, even when individual changes are substantial.

I use the mean absolute subject specific difference to measure the magnitude of condition effects across participants, without requiring them to have the same direction. I then use trial level condition label permutations within each participant to build a null distribution and assess whether the observed magnitude is larger than expected by chance.

An important point is that zero is not the expected value of the mean absolute difference under the null hypothesis. Even without a true condition effect, random trial variation produces nonzero subject specific differences, which become positive when taking their absolute values. Therefore, the observed statistic is compared with a permutation based null distribution rather than directly with zero.

The examples in the notes focus on power, but the approach may also make sense for other EEG measures, particularly peak frequency. It may be more directly relevant to peak frequency (or other representative features of frequency) than to power, but I do not make this distinction in the method (or, more generally, in my perspective, although I can understand why people might do so).