Workshop / Overview

Material used during the workshop is available here:
https://github.com/fma0/AMLD

As the amount and complexity of data collected in neuroscience increases, advanced algorithmic techniques have become imperative for analysing them. In particular, deep learning is often used in basic and clinical neuroscience. Algorithmic decisions have the potential to assist clinical decision making, and to revolutionize basic research, by identifying patterns in rich neurological datasets. 

While the use of these algorithms has an enormous potential for bringing positive change, this comes with a huge cost, as algorithms are often treated as a black box. Oftentimes, it is obscure which features of a dataset an algorithm is using to make a decision, raising questions about interpretability. Moreover, if the training data are misrepresentative of the population variability, AI is prone to reinforcing biases, which can lead to misdiagnosis, or misleading results [1]. 

In this workshop we will explore how AI algorithms can be applied on EEG, a non-invasive technique for measuring neural activity that is commonly used in research and clinics as a diagnostic tool. We will compare ‘traditional’ machine learning algorithms, like linear classifiers, to convolutional neural networks (CNNs), in terms of training and classification performance. Then, we will focus on questions of feature interpretability and algorithmic bias. Attendees will be divided into small groups, where they will have the opportunity to brainstorm potential pitfalls and ways to improve algorithmic interpretability, and reduce bias in neuroscience and EEG data.

Attendees will then have the opportunity to extract and interpret features that CNNs are using to make an algorithmic decision [2], and they will learn how these features can be linked to functions of the human brain. Last, we will provide examples of bias in the context of EEG data (e.g. due to unbalanced datasets), and ways to deal with bias, such as data augmentation techniques.  

Overall, attendees will acquire a theoretical understanding and hands-on experience on (a) training and testing CNNs for EEG data; (b) extracting and exploring features of the data that were important for the network’s decisions; (c) detecting and dealing with bias in the data.

All exercises will be completed in the form of jupyter-notebooks, which attendees can then reuse after the workshop.  

References:

  • [1] Norori N, Hu Q, Aellen F, Faraci F, Tzovara A (2021). Addressing bias in big data and AI for health care: A call for open science, Patterns.
  • [2] Aellen FM, Kavis-Göktepe P, Apostolopoulos S, Tzovara A (2021). Convolutional neural networks for decoding electroencephalography responses and visualizing trial by trial changes in discriminant features, Journal of Neuroscience Methods.

Workshop / Outcome

Attendees will gain insights on how AI can be used in the field of neuroscience for clinical and research applications. They will gain practical skills and hands-on experience with machine learning algorithms for analyzing neuroscience data, and will learn to train CNNs on exemplar datasets. Moreover, attendees will explore ways to extract and interpret features of the classifiers with respect to their relevance for brain functions, and possible ways to deal with biased datasets.

Workshop / Difficulty

Intermediate level

Workshop / Prerequisites

Participants will need a laptop with a browser and internet connection, we will run python jupiter notebooks, either locally or if not possible over google Colab. Participants should have some python programming skills (intermediate). Any further information can be found on: https://neuro.inf.unibe.ch/menu/open.html


Track / Co-organizers

Athina Tzovara

Assistant Professor, University of Bern

Florence Aellen

PhD Candidate, University of Bern

AMLD EPFL 2022 / Workshops

MLOps on AWS: a Hands-On Tutorial

With Gabriele Mazzola, Emanuele Fabbiani, Marco Paruscio, Matteo Moroni, Marta Peroni & Gabriele Orlandi

09:00-13:00 March 262ABC

Close the Gap between Proof-of-Concept and Data Science Product

With Dimira Petrova, Antoni Ivanov & Dako Dakov

10:00-16:00 March 263BC

Designing Effective Visualisations to Communicate Data Stories

With Jacqueline Stählin, Charlotte Cabane, Diana Mitache & Sebastian Baumhauer

10:00-16:00 March 264ABC

AMLD / Global partners