Chaire ANR IA: BrAIN Bridging Artificial Intelligence and - - PowerPoint PPT Presentation

▶
chaire anr ia brain bridging artificial intelligence and
SMART_READER_LITE
LIVE PREVIEW

Chaire ANR IA: BrAIN Bridging Artificial Intelligence and - - PowerPoint PPT Presentation

Sep 25, 2023 429 likes •711 views

Chaire ANR IA: BrAIN Bridging Artificial Intelligence and Neuroscience Alexandre Gramfort alexandre.gramfort@inria.fr INRIA, Universit Paris-Saclay CEA Neurospin Sept. 2020 Supervised learning with fMRI y X Image, sound, task

slide-1
SLIDE 1

Chaire ANR IA: “BrAIN” Bridging Artificial Intelligence and Neuroscience

  • Sept. 2020

Alexandre Gramfort alexandre.gramfort@inria.fr INRIA, Université Paris-Saclay CEA Neurospin

slide-2
SLIDE 2

Alexandre Gramfort Chaire IA BrAIN

Supervised learning with fMRI

2

Image, sound, task

fMRI volume

Scanning Decoding

Objective: Predict y given X or learn a function

s t i m

Any variable: healthy?

y

  • X

f : X → y

slide-3
SLIDE 3

https://paris-saclay-cds.github.io/autism_challenge/

Precision medicine / Biomarkers

slide-4
SLIDE 4

Alexandre Gramfort Chaire IA BrAIN

Why more data is better?

4

Figure from [Gramfort et al. 2011]

  • 5 subjects
  • 12 sessions (more than

1000 scans)

  • Binary classification (face
  • vs. house)
  • Test of 2 left-out sessions

Data from [Haxby et al. 2001]

slide-5
SLIDE 5

Alexandre Gramfort Chaire IA BrAIN

Why more data is better?

4

Figure from [Gramfort et al. 2011]

  • 5 subjects
  • 12 sessions (more than

1000 scans)

  • Binary classification (face
  • vs. house)
  • Test of 2 left-out sessions

Data from [Haxby et al. 2001]

The more data the better

slide-6
SLIDE 6

Alexandre Gramfort Chaire IA BrAIN

Why more data is better?

4

Figure from [Gramfort et al. 2011]

  • 5 subjects
  • 12 sessions (more than

1000 scans)

  • Binary classification (face
  • vs. house)
  • Test of 2 left-out sessions

Data from [Haxby et al. 2001]

The more data the better Almost 100% (no noise)

slide-7
SLIDE 7

Problem: “big data” in science is generally unsupervised

slide-8
SLIDE 8

Project 1

  • Objective: Learning representations from neural time

series with self-supervision and data augmentation

slide-9
SLIDE 9

Project 1

  • Objective: Learning representations from neural time

series with self-supervision and data augmentation

Self-supervision to the rescue

E.g.: Jigsaw puzzle task from Noroozi & Favaro (2016) Other examples: word2vec, BERT, nonlinear ICA, etc. In a nutshell: use the structure of the data to pretrain a feature extractor with a supervised (“pretext”) task – then use the features.

Original image Input patches Output

slide-10
SLIDE 10

Project 1

  • Objective: Learning representations from neural time

series with self-supervision and data augmentation

[Banville et al. MLSP 2019]

SSL to learn on sleep EEG

slide-11
SLIDE 11

Problem: What pretext task makes sense for EEG/ MEG?

  • Use knowledge about sleep (slow cycles)
  • Theoretical approaches based on recent

results on identifiability of non-linear ICA

slide-12
SLIDE 12

Alexandre Gramfort Chaire IA BrAIN

Possible Self Sup. Tasks

9

Time (e.g., minutes, hours) Amplitude (e.g., V)

ch1 ch2 ch3 ch4

Relative positioning (RP)

Logistic regression

  • Sampling

Training 1 2

Predict if 2 windows of data are close in time Other approaches: CPC [Oord et al. 2018], PCL [Hyvärinen et al. 2017] etc.

slide-13
SLIDE 13

Alexandre Gramfort Chaire IA BrAIN

Project 1

10

  • Objective: Learning representations from neural time

series with self-supervision and data augmentation

slide-14
SLIDE 14

Problem: Augmenting MEG/EEG data is not as simple as for images or speech

  • Use the physics of MEG/EEG
  • Use knowledge/availability of pure noise
  • Use knowledge about neuroscience (freq.

shifts, biophysiological models)

slide-15
SLIDE 15

Problem: Augmenting MEG/EEG data is not as simple as for images or speech

  • Use the physics of MEG/EEG
  • Use knowledge/availability of pure noise
  • Use knowledge about neuroscience (freq.

shifts, biophysiological models)

We want to learn how to augment neuroscience data!

slide-16
SLIDE 16

Alexandre Gramfort Chaire IA BrAIN

Problem of dataset variability

12

  • ≠ recording devices / scanners
  • ≠ EEG channels / fMRI sequence parameters
  • ≠ preprocessing steps
  • ≠ populations: ages, sexes, clinical disorders…
  • ≠ labeling guidelines
slide-17
SLIDE 17

Alexandre Gramfort Chaire IA BrAIN

Problem of dataset variability

12

  • ≠ recording devices / scanners
  • ≠ EEG channels / fMRI sequence parameters
  • ≠ preprocessing steps
  • ≠ populations: ages, sexes, clinical disorders…
  • ≠ labeling guidelines
  • Pooling datasets to increase n can reduce performance
  • Performance on new dataset can drop
slide-18
SLIDE 18

Alexandre Gramfort Chaire IA BrAIN

Problem of dataset variability

12

  • ≠ recording devices / scanners
  • ≠ EEG channels / fMRI sequence parameters
  • ≠ preprocessing steps
  • ≠ populations: ages, sexes, clinical disorders…
  • ≠ labeling guidelines
  • Pooling datasets to increase n can reduce performance
  • Performance on new dataset can drop

[Torralba and Efros, 2011]

slide-19
SLIDE 19

Alexandre Gramfort Chaire IA BrAIN

Domain adaptation with EEG sleep

13

[Chambon et al., Domain adaptation with optimal transport improves EEG sleep stage classifiers, PRNI 2018]

  • Train dataset: MESA [Dean et al. 2016]
  • Test dataset: MASS-session 3 [O’Reilly et al. 2014]
  • 3 EEG + 2 EOG channels
slide-20
SLIDE 20

Alexandre Gramfort Chaire IA BrAIN

Domain adaptation with EEG sleep

13

[Chambon et al., Domain adaptation with optimal transport improves EEG sleep stage classifiers, PRNI 2018]

  • Train dataset: MESA [Dean et al. 2016]
  • Test dataset: MASS-session 3 [O’Reilly et al. 2014]
  • 3 EEG + 2 EOG channels

Domain adaptation improves performance

slide-21
SLIDE 21

How do we impact neuroscience and medicine?

slide-22
SLIDE 22

Predict of brain “fragility” for optimal drug dosage across age

Joint work with:

slide-23
SLIDE 23

MNE software for processing MEG and EEG data, A. Gramfort, M. Luessi, E. Larson, D. Engemann, D. Strohmeier, C. Brodbeck, L. Parkkonen, M. Hämäläinen, Neuroimage 2013

https://mne.tools/

Transfer + impact with MNE

slide-24
SLIDE 24

Alexandre Gramfort Chaire IA BrAIN

Objectives

17

  • O1. Learn with no-supervision on noisy and complex

multivariate signals

  • O2. Learn end-to-end predictive systems from limited

data exploiting physical constraints

  • O3. Learn from data coming from many different source

domains

  • O4. Develop high-quality software tools that can reach

clinical research BrAIN objective: Develop the next ML paradigms to extract knowledge from physiological signals

slide-25
SLIDE 25

Alexandre Gramfort Chaire IA BrAIN

Team

18

  • Denis Engemann
  • Thomas Moreau
  • 1 Post-doc
  • 1 Engineer
  • 3 PhDs
  • INSERM team at Larib. for clinical cases
  • Aapo Hyvärinen as external collaborator/visitor
slide-26
SLIDE 26

GitHub : @agramfort Twitter : @agramfort

http://alexandre.gramfort.net

Contact

"An approximate answer to the right problem is worth a good deal more than an exact answer to an approximate problem. ~ John Tukey"