Titulo Estágio
Artificially augmenting neural signals for Brain-Computer Interfaces
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
CISUC
Enquadramento
A Brain-Computer Interface (BCI) is a system that provides a direct communication between the brain and a computer or external device. This system interprets the brain activity via neural decoders and translates it into commands that can be used to control devices or programs.
Different types of neuroimaging modalities can be used to implement a BCI. Due to the low risk, low cost, and convenience, EEG-based non-invasive BCIs are the most popular BCIs. Designing an accurate EEG decoder is crucial and challenging, e.g., due to the low signal-to-noise ratio of EEG.
Deep neural networks (DNNs), originally proposed in the field of image, speech and natural text processing, are emerging as powerful EEG decoders. With these models, the most relevant features for neural decoding can be automatically extracted from the input signals via the sequence of non-linear transformations based on hierarchical representations. Furthermore, input signals can be raw/lightly pre-processed, thus, they preserve most of the information contained in the EEG for the decoding task. However, these algorithms generally introduce many trainable parameters (e.g., >5K parameters), thus, require large datasets to achieve higher decoding performance than traditional decoders.
Collecting large-scale EEG signals is difficult, as EEG signals are highly susceptible to changes in physiological and psychological conditions, causing high variability across subjects and sessions. This may limit the decoding capabilities of deep learning approaches. One solution that is growing interest in the state-of-the-art is the adoption of data augmentation approaches to artificially increase a compact and high-quality set of signals [1]. Among these techniques, simple procedures such as applying flip, crop, rotation transformations were used, in addition to noise injection (in spatial, temporal, and frequency domains). Furthermore, deep neural networks were also proposed as data augmenters (e.g., generative adversarial networks).
Despite the raising of these promising techniques, it is not clear whether data augmentation could always be beneficial for DNNs in EEG-based BCIs and what is the optimal data augmentation strategy.
Objetivo
Investigate and implement the main data augmentation strategies of the state-of-the-art for EEG-based BCIs, in particular to facilitate the use of DNNs in BCIs. This includes:
- Identify the most relevant and complete datasets to test data augmenters
- Identify the state-of-the-art DNNs and data augmenters
- Train and evaluate (offline) DNN-based decoders with or without the augmenters to evaluate empirically their beneficial effect in BCIs
- Report and analyze the results
Plano de Trabalhos - Semestre 1
1. State-of-the-art review: DNNs and data augmenters for EEG signals
2. Identification and familiarization with datasets to use
3. Definition of the methodological approach
4. Writing of the Intermediate Report
Plano de Trabalhos - Semestre 2
5. Research and implementation (based on the literature) of DNNs and data augmentation strategies
6. Experimental analysis and reporting of the developed models with/without augmenters
7. Writing of a scientific paper and dissertation
Condições
Access to a GPU-enabled mainframe for running DNNs
Good programming skills (Matlab and/or Python)
Possibility of partial funding through a scholarship depending on the curriculum of the candidate.
Observações
[1] Elnaz Lashgari, Dehua Liang, Uri Maoz, (2020) Data augmentation for deep-learning-based electroencephalography. Journal of Neuroscience Methods, Volume 346, 108885, ISSN 0165-0270,
doi: 10.1016/j.jneumeth.2020.108885. (https://www.sciencedirect.com/science/article/pii/S0165027020303083)
Co-suppervision by Davide Borra, University of Bologna
Orientador
Marco António Machado Simões
msimoes@dei.uc.pt 📩