Joseph retweeted
โจ๐ฐpreprint - From prediction to prescription: Machine learning and Causal Inference
hal.science/hal-04774700
Machine learning for individual treatment effect
Didactic and precise, bridging statistical, computational, and epidemiological thinking.
#statstwitter #epitwitter
1/5
Joseph retweeted
Would you have thought that object recognition CNNs can encode information on object number across changes in identity and context? New preprint, with Thomas Chapalain and @BertrandThirion, on visual #number_sense: biorxiv.org/content/10.1101/โฆ
Joseph retweeted
๐ง Introducing GREEN (Gabor Riemann EEG Net), a new modular neural network architecture for #EEG analysis. GREEN combines the power of wavelet transforms ใฐ๏ธ and Riemannian geometry ๐ to learn sparse and interpretable representations of EEG data. It's designed with modularity in mind, making it easy to adapt and extend for different research questions.
One of GREEN's distinct design features is its ability to learn task-specific wavelets, which could help advance our understanding of brain function and help identify biomarkers for neurological and psychiatric disorders. ๐ฉบ By bridging the gap between deep learning ๐ค and traditional EEG analysis techniques ๐ง ใฐ๏ธ๏ธ, GREEN opens up new avenues for exploring the brain's complex electrical signals. โก๏ธ
Explore our preprint to discover how GREEN's modular design, interpretable representations, and learnable wavelets can advance #neuroscience and #biomedical research: biorxiv.org/content/10.1101/โฆ #deeplearning #opensource #EEG
Joint work with @JP4illard & @hippneuro !
Joseph retweeted
To preprocess or not to preprocess your #EEG (when building #ML models) ๐ฎ? ๐ซWe are thrilled to share our latest #preprint, studying the challenge of learning brain-specific biomarkers from EEG ๐ง ๐ถ using #ML โ๏ธ๐ฅ๏ธ: biorxiv.org/content/10.1101/โฆ
We compile arguments and evidence from benchmarking age- & sex-prediction on > 2600 EEGs from two large public datasets.
We found that basic artifact rejection consistently led to better model performance, whereas removal of ocular and muscle artifacts hampered performance. As it turns out that those peripheral signals are predictive themselves!
Our results therefore argue in favor of the need to diligently process EEG data, if the goal is to have brain-specific biomarkers (and if prediction is not the only objective).
Our efforts to build more interpretable #ML models for EEG led us to extending the established Morlet wavelet methodology for spectral analysis of EEG to accommodate state-of-the-art ML models based on covariance matrices. This allowed us to perform head-to-head comparisons between classical EEG features and frequency-specific model predictions for, both, brain and artifact signals.
Compared to classical band-pass filtering, wavelets even led to improvements in prediction performance.
Joint work with Philipp Bomatter, @JP4illard, Pilar Garces & Jรถrg F Hipp.

ALT Conceptual framework for biomarker learning from EEG (a). Morelet wavelets provide a unified representation for computing classical spectral features (i.e. log power) and covariances used by advanced ML models for EEG (b). Modeling and benchmarking setup for studying the impact of peripheral EEG generators on ML model predictions (c).
Joseph retweeted
Our paper on data augmentation for EEG data was accepted by the Journal of Neural Engineering!
It all started w the nice work done by @JP4illard during his internship @Parietal_INRIA! Congrats to him & our amazing co-authors @tomamoral & @agramfort!
Short ๐งต on our findings ๐!
Joseph retweeted
Ever wondered how to carry data augmentation depending on the data label?
That's what we've investigated in our #iclr2022 paper w/ @tomamoral, J. Paillard & @agramfort
CADDA: Class-wise Automatic Differentiable Data Augmentation for EEG signals
๐ arxiv.org/abs/2106.13695
๐งต1/6
Red Hot ๐ถ๏ธ #ASMST โ at Clermont-Ferrand, France














