Bridging Artificial Intelligence and Neuroscience – BrAIN

After pioneering breakthroughs in computer vision, speech processing or natural language processing, ML has now to face new challenges in order to impact various scientific disciplines and in particular health related applications. When considering medical applications, statistical and computational problems emerge. i) The first problem is related to the absence or limited amount supervision for algorithms: supervised predictive models need so-called annotations or labels to be trained and tested, and unfortunately too few medical applications can provide enough of these. ii) The second problem is related to what can be phrased as dataset variability, or in more statistical terms, distribution or covariate shifts. What has been called in computer vision the “dataset bias” problem, implies that training on data from a certain hospital is likely to provide less powerful prediction when testing on data from a different hospital. iii) The third problem is related to the difficulty of bringing the state-of-the-art tools to an environment that is not dominated by computer scientists but biologists, neuroscientists, psychologists, medical doctors. BrAIN aims to provide the next generation of ML models and algorithms for efficient statistical learning in the absence of strong labels and large sample sizes. BrAIN will leverage clear use-cases in clinical and cognitive neuroscience (anaesthesia, disorders of consciousness, sleep medicine) to address general ML challenges: 1) study of various self-supervised learning tasks to learn from long and noisy temporal data 2) learning to augment data and increase sample sizes 3) robust learning in the presence of distribution shifts 4) development of tractable algorithms easy to use by non-experts.

Benchopt: Reproducible, efficient and collaborative optimization benchmarks
Moreau T., Massias M., Gramfort A., Ablin P., Charlier P., Dagréou M., la Tour T., Durif G., Dantas C., Klopfenstein Q., Larsson J., Lai E., Lefort T., Malézieux B., Moufad B., Nguyen B., Rakotomamonjy A., Ramzi Z., Salmon J., Vaiter S. (2022)

Repurposing EEG monitoring of general anaesthesia for building biomarkers of brain ageing: An exploratory study
Sabbagh D., Cartailler J., Touchard C., Joachim J., Mebazaa A., Vallée F., Gayat {., Gramfort A., Engemann D. (2022)
medRxiv .

The Optimal Noise in Noise-Contrastive Learning Is Not What You Think
Chehab O., Gramfort A., Hyvarinen A. (2022)
The 38th Conference on Uncertainty in Artificial Intelligence

Robust learning from corrupted EEG with dynamic spatial filtering
Banville H., Wood S., Aimone C., Engemann D., Gramfort A. (2022)
NeuroImage 251: (118994).

A reusable benchmark of brain-age prediction from M/EEG resting-state signals
Engemann D., Mellot A., Höchenberger R., Banville H., Sabbagh D., Gemein L., Ball T., Gramfort A. (2021)
bioRxiv .

Shared Independent Component Analysis for Multi-Subject Neuroimaging
Richard H., Ablin P., Thirion B., Gramfort A., Hyvärinen A. (2021)
Advances in Neural Information Processing Systems 34 (NeurIPS)

Uncovering the structure of clinical EEG signals with self-supervised learning
Banville H., Chehab O., Hyvärinen A., Engemann D., Gramfort A. (2021)
Journal of Neural Engineering 18: (046020).

Deep Recurrent Encoder: A scalable end-to-end network to model brain signals
Chehab O., Defossez A., Loiseau J., Gramfort A., King J. (2021)

CADDA: Class-wise Automatic Differentiable Data Augmentation for EEG Signals
Rommel C., Moreau T., Paillard J., Gramfort A. (2022)
International Conference on Learning Representations

HNPE: Leveraging Global Parameters for Neural Posterior Estimation
Rodrigues P., Moreau T., Louppe G., Gramfort A. (2021)
Advances in Neural Information Processing Systems 34 (NeurIPS)

Push the limits of machine learning on neuroscience signals.

Voir alexandre.gramfort.net/anr-brain.html

Artificial intelligence (AI) with recent progress in statistical machine
learning (ML) is currently aiming to revolutionise how experimental science is
conducted. In physics, chemistry, biology, neuroscience or medicine,
data is now the driver of new theoretical insights and new scientific hypotheses.
Supervised learning and predictive models are now used to
assess if something is ``predictable'': Can I predict what people
``think'' from neural signals? Can I predict from DNA if a patient will suffer
from cancer? ML is now used as a
replacement for classical statistical hypothesis testing.
In healthcare, one talks about precision medicine, virtual patients
with the vision that artificial intelligence will allow to
have individualised predictions from genomic, physiological or imaging data.

After pioneering breakthroughs in computer vision,
speech processing or natural language processing, ML has now
to face new challenges in order to impact various scientific
disciplines and in particular health related applications.
When considering medical applications, statistical and computational
problems emerge.
i) The first problem is related to the absence or limited amount
supervision for algorithms: supervised predictive models need
so-called annotations or labels to be trained and tested,
and unfortunately too few medical applications can provide enough of these.
ii) The second problem is related to what can be phrased as dataset
variability, or in more statistical terms, distribution
or covariate shifts. What has been called in computer vision
the ``dataset bias'' problem, implies that training on data from
a certain hospital is likely to provide less powerful
prediction when testing on data from a different hospital.
iii) The third problem is related to the difficulty of bringing
the state-of-the-art tools to an environment that is not
dominated by computer scientists but biologists, neuroscientists,
psychologists, medical doctors.

BrAIN will provide the next generation of ML models and algorithms
for efficient statistical learning in the absence of
strong labels and large sample sizes.
BrAIN will leverage clear use-cases in clinical and cognitive
neuroscience (anaesthesia, disorders of consciousness, sleep medicine)
to address general ML challenges:
1) study of various self-supervised learning tasks to learn from long and noisy temporal data
2) learning to augment data and increase sample sizes
3) robust learning in the presence of distribution shifts
4) development of tractable algorithms easy to use by non-experts.