Channel set invariance for neural networks

Problem

Despite the well-established standards for EEG electrode layout like the 10-10 montage, EEG recordings obtained in different labs or across studies are not straightforward to compare. There are usually important differences between the datasets, but also within them. These differences can be:

• the use of different channel sets or sensor placement routines
• different noise distributions, or changes of the noise structure over time
• temporarily unusable channels (electrical connection lost, faulty wire, …)

For this reason, spatial filters can hardly be re-used between sessions, let alone datasets, which is an obstacle for transfer learning approaches in EEG decoding problems. The existing solutions to do transfer learning with heterogeneous channel sets are:

• train a model on the common channel subset → restrictive, if a new recording has a faulty channel, as the whole model has to be re-trained without this channel
• interpolate the missing channels → obtained full channel set may not be full rank any more
• source reconstruction techniques → assumptions need to be made for source reconstruction, their choice is non-trivial

Objective

The objective of this project is to help in the development of a neural network architecture invariant to the set of channels used. Such an architecture would allow to realize an end-to-end learning approach across multiple datasets with heterogeneous channel sets. Ideally, this architecture would be able to:

• handle EEG signals containing an arbitrary number of channels
• produce results independently of the channels used
• reach the same performances as non-channel independent architectures (like EEGNet) when the channel set is fixed / complete
• seamlessly ignore corrupted channels, and degrade gracefully
• cope in real time with potential noise distribution shifts

Skills required

• Good programming skills in Python
• Experience with the pytorch library
• Optional: experience with deploying pipelines on GPU clusters
• Optional: experience with EEG/BCI data