Diffusion Modelling approaches to EEG-based Auditory Attention Decoding
(2024)Department of Automatic Control
- Abstract
- Machine learning models can analyze physiological data, such as electroencephalography (EEG), for various classification tasks. One such task is Auditory Attention Decoding (AAD), aimed at identifying the sound a person is actively attending to, offering significant benefits for users of hearing aids. However, EEG data often exhibits a low signal-to-noise ratio, and its collection is often expensive, cumbersome and requires trained specialists. Additionally, gathering EEG data over prolonged periods of time presents challenges. The resulting scarcity of available EEG data to train models can be alleviated by using generative models, which generate new examples from the data they were trained on.
Diffusion Probabilistic Models (DPMs) have... (More) - Machine learning models can analyze physiological data, such as electroencephalography (EEG), for various classification tasks. One such task is Auditory Attention Decoding (AAD), aimed at identifying the sound a person is actively attending to, offering significant benefits for users of hearing aids. However, EEG data often exhibits a low signal-to-noise ratio, and its collection is often expensive, cumbersome and requires trained specialists. Additionally, gathering EEG data over prolonged periods of time presents challenges. The resulting scarcity of available EEG data to train models can be alleviated by using generative models, which generate new examples from the data they were trained on.
Diffusion Probabilistic Models (DPMs) have in recent years emerged as the state-of-the-art of generative models within the image domain, showing widespread success in models such as Stable Diffusion and DALL-E. This work investigates whether this success can extend to the domain of multichannel time series data, specifically EEG data. The diffusion models were trained on 1-second EEG data segments and were used as a data augmentation tool for 3 different classification tasks, including AAD. Our findings indicate that diffusion models can effectively generate realistic EEG data, supported by both a visual comparison and a measure of Jensen-Shannon divergence to the real EEG data distribution. In addition to this, a significant improvement in mean performance was achieved in our Locus of Attention (LoA) task, where we classify between a test subject attending to a left or right speaker. Here an approximate classification accuracy of 71% was achieved compared to our baseline of 70.4%. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9150900
- author
- Rannaleet, David and Gunnarsson, Victor
- supervisor
- organization
- year
- 2024
- type
- H3 - Professional qualifications (4 Years - )
- subject
- report number
- TFRT-6227
- other publication id
- 0280-5316
- language
- English
- id
- 9150900
- date added to LUP
- 2024-04-17 14:03:02
- date last changed
- 2024-04-17 14:03:02
@misc{9150900,
abstract = {{Machine learning models can analyze physiological data, such as electroencephalography (EEG), for various classification tasks. One such task is Auditory Attention Decoding (AAD), aimed at identifying the sound a person is actively attending to, offering significant benefits for users of hearing aids. However, EEG data often exhibits a low signal-to-noise ratio, and its collection is often expensive, cumbersome and requires trained specialists. Additionally, gathering EEG data over prolonged periods of time presents challenges. The resulting scarcity of available EEG data to train models can be alleviated by using generative models, which generate new examples from the data they were trained on.
Diffusion Probabilistic Models (DPMs) have in recent years emerged as the state-of-the-art of generative models within the image domain, showing widespread success in models such as Stable Diffusion and DALL-E. This work investigates whether this success can extend to the domain of multichannel time series data, specifically EEG data. The diffusion models were trained on 1-second EEG data segments and were used as a data augmentation tool for 3 different classification tasks, including AAD. Our findings indicate that diffusion models can effectively generate realistic EEG data, supported by both a visual comparison and a measure of Jensen-Shannon divergence to the real EEG data distribution. In addition to this, a significant improvement in mean performance was achieved in our Locus of Attention (LoA) task, where we classify between a test subject attending to a left or right speaker. Here an approximate classification accuracy of 71% was achieved compared to our baseline of 70.4%.}},
author = {{Rannaleet, David and Gunnarsson, Victor}},
language = {{eng}},
note = {{Student Paper}},
title = {{Diffusion Modelling approaches to EEG-based Auditory Attention Decoding}},
year = {{2024}},
}