イベント情報
【GIR公開セミナー】Dr. Andrzej Cichocki / Polish Academy of Science (Poland), Dr. Fabien Lotte / 国立情報学自動制御研究所 (INRIA)（フランス）, Dr. Sébastien Rimbert / Inria Bordeaux Sud-Ouest（フランス）
| 日時 | 2025.3.5（15：30～18：00） |
|---|---|
| 会場 |
ミーティングID：841 2225 5201 パスコード：787128 |
| 講演者 | Dr. Andrzej Cichocki, ポーランド科学アカデミー （ポーランド） Dr. Fabien Lotte, 国立情報学自動制御研究所 (INRIA)（フランス） Dr. Sébastien Rimbert, Inria Bordeaux Sud-Ouest（フランス） |
| 講演タイトル | ◆Dr. Andrzej Cichocki "Deep Neural Networks for BCI: Recent trends and innovations" ＜要旨＞ In this talk we briefly review various architectures and associated learning algorithms of deep neural networks (DNN) for Brain Computer Interfaces (BCI), especially for classification, recognition and clustering tasks. We mostly focus on novel multiscale DNN, which exploits not only temporal but also spatio-temporal information in EEG signals and employs various kinds of sophisticated convolutions. Several modules/layers like Squeeze and Excitation (SE) modules and improved attention blocks in the Transformer, Dynamic convolutional layer and selected ML methods feature extraction and selection will be discussed. Moreover, the Lightweight deep neural networks for BCI and efficient compression of large scale multilayer DNNs via tensor networks will also be presented. ◆Dr. Fabien Lotte "Machine Learning for EEG-based BCIs: recent advances and ambushes" ＜要旨＞ Brain-Computer Interfaces (BCIs) are systems that translates users' brain activity, typically measured using ElectroEncephaloGraphy (EEG), into commands for an application. Whereas BCIs are very promising for various applications, e.g., brain-based wheelchair control or plane pilots' mental state monitoring, they are not reliable, and not so usable. They indeed make frequent mistakes in recognizing the users' mental commands, their performances are usually very variable between users, and even within users, e.g., across time, and they require lengthy and tedious calibration and re-calibration over time. Considerable research efforts are being conducted to address these issues thanks to dedicated EEG machine learning algorithms. Two main families of promising algorithms stand at the forefront of these efforts: Riemannian geometry and Deep Learning algorithms. Riemannian geometry is arguably the current state-of-the-art for user-specific EEG classifier design, while deep learning seems incredibly more promising for cross-users/cross-data set EEG classifier design. These increasing research efforts lead to some advances in the field, but also to new ambushes, that should be avoided. In this talk I will first present some recent advances in Riemannian geometry for EEG-BCI from our group, notably regarding Riemannian methods to handle EEG temporal variability, new Riemannian representations of EEG, methods to visualize the EEG Riemannian manifold and data and method to model distributions of covariance matrices on the Riemannian manifold, we can lead to various new Riemannian classifiers. I will then discuss various ambushes that unfortunately affect more and more BCI studies using machine learning. I will for instance show that simple visual instructions can bias EEG Deep learning models, or that several validation approaches can lead to vastly overestimated classification performances. I will conclude by presenting some recent consensus guidelines to help ensuring Machine Learning research has a positive impact on BCI, while avoiding the many associated ambushes. ◆Dr. Sébastien Rimbert "Enhancing and Predicting MI-BCI Performance with Median Nerve Stimulation" ＜要旨＞ -Context Mental-Imagery-based Brain-Computer Interfaces (MI-BCIs) have advanced significantly, yet challenges remain for real-world applications. Current systems often require extensive training and still struggle with suboptimal accuracy, with 30% of users unable to use them effectively. Additionally, many rely on synchronous interactions, reducing intuitiveness and usability. -Positioning To address these challenges, I propose a new MI-BCI paradigm based on median nerve stimulation (MNS). My research shows that painless MNS during MI tasks enhances BCI classification accuracy. Both MI and MNS induce similar cortical activity changes, and when combined, the ERD/ERS signature becomes more pronounced, improving machine learning classification. Our results show up to 15% accuracy improvement and 20% gains in recall/precision, paving the way for high-performance, asynchronous BCIs. Predicting MI-BCI Performance with Median Nerve Stimulation Beyond improving accuracy, MNS can predict MI-BCI performance. Our studies reveal that the minimum value of ERD induced by MNS exhibits a strong correlation with MI-BCI accuracy (rho = 0.71, p < 0.001). Additionally, this predictor effectively estimates MI-BCI performance with a significant correlation (rho = 0.5, mean absolute error = 9.0, p < 0.01). These findings establish MNS as a reliable, non-invasive, and user-independent predictor of MI-BCI proficiency. By analyzing MNS-induced ERD patterns, we can passively assess a user’s BCI capability and potentially adapt system parameters for an optimized experience or rehabilitation purposes. In this talk, I will present how median nerve stimulation enhances MI-BCI classification and usability while also serving as a predictive tool for user performance. |
| 言語 | 英語 |
| 対象 | どなたでもご参加いただけます。 |
| 共催 | グローバルイノベーション研究院 GRH「動物共生情報学」拠点 卓越大学院プログラム |
| お問い合わせ窓口 | グローバルイノベーション研究院・工学研究院 田中 聡久 e-mail: tanakat (ここに@ を入れてください）cc.tuat.ac.jp |
| 備考 | 本セミナーは、オンライン・対面型の同時開催となります。 |
このページの上部へ