Informatics in the psychological sciences brings fascinating challenges as mental processes or pathologies have fuzzy definition and are hard to quantify. Brain imaging brings rich data on the neural substrate of these concepts, yet it is a non trivial link.
The goal of this presentation is to put forward basic ideas of "psychoinformatics", using advanced processing on brain images to quantify better the elements of psychology.
It discusses how machine learning can bridge brain images to behavior: to describe better mental processes involved in brain activity, or to extract biomarkers of pathologies, individual traits, or cognition.
8. Studying mental processes
- -Results in
a contrast
Study by
oppositions
Results tied to a simple psychological manipulation
bound to a paradigm
G Varoquaux 4
13. Predictive models can broaden theories by
generalizing brain-mind associations to arbitrary new
tasks and stimuli [Varoquaux and Poldrack 2018]
1 Beyond oppositions: encoding
2 Generalizing across tasks
3 Universal cognitive representations
4 Across subjects: biomarkers
G Varoquaux 8
16. 1 Decomposing psychological process
To study vision: Breaking down stimuli
[Hubel and Wiesel 1962]
Neurons receptive to
Gabors (edges)
G Varoquaux 10
17. 1 Decomposing psychological process
To study vision: Breaking down stimuli
[Hubel and Wiesel 1962]
Neurons receptive to
Gabors (edges)
[Logothetis... 1995]
Shapes in inferior
temporal cortex
G Varoquaux 10
18. 1 Decomposing psychological process
To study vision: Breaking down stimuli
Image
V1
cortex
V2
cortex
Inferior
temporal
cortex
Fusiform
face area
Jack?
Is there a “face” region? A “foot” region? A “left big toe” region?
G Varoquaux 10
19. 1 Crafting stimuli for cognitive oppositions
Is there a “face” region? A “foot” region? A “left big toe” region?
vs
G Varoquaux 11
20. 1 Crafting stimuli for cognitive oppositions
Is there a “face” region? A “foot” region? A “left big toe” region?
vs
G Varoquaux 11
21. 1 Crafting stimuli for cognitive oppositions
Is there a “face” region? A “foot” region? A “left big toe” region?
vs
-
G Varoquaux 11
22. 1 Crafting stimuli for cognitive oppositions
Is there a “face” region? A “foot” region? A “left big toe” region?
vs
-
Representing only one aspect of the stimuli:
too much reductionism
G Varoquaux 11
23. 1 Not crafting stimuli for cognitive oppositions
Is there a “face” region? A “foot” region? A “left big toe” region?
vs
-
Representing only one aspect of the stimuli:
too much reductionism
Encoding
Building complex representations of stimuli
Predicting brain response from them
G Varoquaux 11
24. 1 Decomposing visual stimuli
Low-level visual cortex is tuned
to natural image statistics
[Olshausen et al. 1996]
What drives high-level representations?
The concept of “left big toe”?
G Varoquaux 12
25. 1 Decomposing visual stimuli
Low-level visual cortex is tuned
to natural image statistics
[Olshausen et al. 1996]
What drives high-level representations?
Convolutional Net
G Varoquaux 12
26. 1 Brain mapping: encoding with conv nets
StimulirepresentationBrainactivity
Input Layer 1 Layer 5
Linear predictive models
Convolutional Net
G Varoquaux 13
27. 1 Brain mapping: encoding with conv nets
StimulirepresentationBrainactivity
Input Layer 1 Layer 5
Linear predictive models
Convolutional Net
Explains the brain activity from the stimuli
much better than hand-crafted features
On data from [Kay... 2008] natural images
[Huth... 2012] movies
[Eickenberg... 2017]
G Varoquaux 13
28. 1 Brain mapping: encoding with conv nets
[Eickenberg... 2017]
High-level conv-net layer map to high-level visual areas
G Varoquaux 14
29. 1 Brain mapping with conv nets: retinotopy
Mapping response to exentricity: artifical retinotopy
G Varoquaux 15
30. 1 Brain mapping with conv nets: high-level concepts
Opposing face versus place in the [Haxby... 2001] stimuli
G Varoquaux 16
31. [Eickenberg... 2017]
Beyond oppositions: encoding
Recovers the
hierarchy of
visual modules
Generalizes
from natural
images
to retinotopy or
face vs place
G Varoquaux 17
32. 2 Generalizing across tasks
Characterizing the function
of brain structures
How much is observed activity
a consequence of specificities
of the paradigm?
[Schwartz... 2013, Varoquaux... 2018]
G Varoquaux 18
33. 2 Large-scale decoding for reverse inference
Cat
Dog
Ouaf
Miaou
[Poldrack... 2009]
G Varoquaux 19
34. 2 Large-scale decoding for reverse inference
Cat
Dog
Ouaf
Miaou
[Poldrack... 2009]
Variability: an opportunity and a challenge
Technical heterogeneity (scanner, stimulus modality...)
Paradigmatic isolation (cognition) [Newell 1973]
G Varoquaux 19
35. 2 Generalizing to arbitrary paradigms
[Schwartz... 2013, Varoquaux... 2018]
What is this
brain doing?
Describe tasks by their cognitive components
Multi-label prediction: presence or absence of each label
Visual
Auditory
Foot
Hand
Calculation
Reading
Checkboard
Face
Place
Object
Digit
Saccade
...
...
...
...
...
...
G Varoquaux 20
36. 2 Generalizing to arbitrary paradigms
[Schwartz... 2013, Varoquaux... 2018]
What is this
brain doing?
Describe tasks by their cognitive components
Multi-label prediction: presence or absence of each label
Visual
Auditory
Checkboard
Face
...
...
...
...
Prediction across studies
Describe a task never seen
from the brain activity it evokes
G Varoquaux 20
37. 2 Generalizing to arbitrary paradigms
[Schwartz... 2013, Varoquaux... 2018]
What is this
brain doing?
+
+
Regions specific to that predict
facets of cognition
G Varoquaux 20
38. 2 Decoding mega-analysis
[Schwartz... 2013, Varoquaux... 2018]
30 studies 837 subjects 196
experimental
conditions
6919 activation maps
Different labs
Dehaene
Poldrack
Wager
...
Various cognitive domains
Language
Vision
Decision making
Mathematics
...
All manually preprocessed, labeled, and curated
G Varoquaux 21
39. 2 Decoding mega-analysis
[Schwartz... 2013, Varoquaux... 2018]
30 studies 837 subjects 196
experimental
conditions
6919 activation maps
Different labs
Dehaene
Poldrack
Wager
...
Various cognitive domains
Language
Vision
Decision making
Mathematics
...
All manually preprocessed, labeled, and curated
Decoding arbitrary new paradigms
visualauditoryVcheckerboard
Hcheckerboardobjectsscramblefacesplacesdigits
wordsrighthandlefthandrightfootleftfootlanguagehumansound
non-humansound
0.2
0.4
0.6
0.8
1.0
Predictionscore
(ROCAUC)
ontologydecoder
logisticregression
naiveBayes
neurosynth
0.3
0.2
0.1
0.0
+0.1
+0.2
+0.3
Relative score
G Varoquaux 21
40. 2 Mapping: regions specific to facets of cognition
[Schwartz... 2013, Varoquaux... 2018]
Contrasts: good rejection of confounds
Decoding: reverse inference
⇒ Consensus to define regions
The “place” concept
Forward
term effect
Brain response to tasks with
the “place” concept
Forward
ontology contrast
Contrasting responses to
“place” and related concepts
Reverse
decoder map
Regions that predict a place-
recognition task
Consensus
forward & reverse
Regions selected by forward
and reverse inference
Replaces the careful crafting of control conditions
G Varoquaux 22
41. 2 A functional atlas: central gyrus, motor networks
[Schwartz... 2013, Varoquaux... 2018]
Neurosynth
Reverse inference
(Logistic regression)
Decoding in an
ontology
z = -15
Reverse inference
-15
nference DecodDecoding in an
visual faces places objects scrambled words digits horizontal - vertical
checkerboard
Visual regions
Contrasts
-based
analysis
Neurosynth
reverse
inference
Decoding
and
ontology
+ Language and auditory mapping in temporal cortex
+ Calculation and spatial attenation in IPS
+ Mapping the motor system
G Varoquaux 23
42. [Varoquaux... 2018]
Generalizing across tasks
Cat
Dog
Ouaf
Miaou
Multi-label
prediction
Prediction
(discriminative
models) more robust
to heterogeneity
Decoding
⇒ evidence beyond
⇒ a paradigm
Atlasing cognition
G Varoquaux 24
43. Need an ontology of cognition
Episodic memory
working memory
lexical memory
Syntax comprehension
Reading
Mental calculation
Shape recognition
Face recognition
Memory
Language
Vision
G Varoquaux 25
45. 3 Universal cognitive
representations [Mensch... 2017]
Face Shape
Reading Shape
Language Music
Saccades CalculationMulti-task learning
Solve many related but different problems
Learn commonalities
intermediate representations
G Varoquaux 26
47. 3 Deep architecture
[Bengio 2009]
...
Great for multiple output (tasks)
Millions of parameters, thousands of data points
G Varoquaux 27
48. 3 Deep architecture
2
2
Great for multiple output (tasks)
Millions of parameters, thousands of data points
Simplify
G Varoquaux 27
49. 3 Shallow architecture
2
linear model
Great for multiple output (tasks)
Millions of parameters, thousands of data points
Simplify simplify more
G Varoquaux 27
50. 3 Shallow architecture
2
2
Great for multiple output (tasks)
Millions of parameters, thousands of data points
Simplify simplify more
G Varoquaux 27
51. 3 Deep linear model – multi-task – Bayesian
4TB resting-state data
HCP900
OpenfMRI
HCP
Camcan
Brainomics
42000 task
fMRI contrast maps
for one model
...
Many
rest & task
fMRI studies
...
x
Information flow
x Learned network
combination
1st layer
assignment
Non-negative
matrix factorization
SUCCESSFUL
GO
HOUSE
HORIZONTAL
CHECKERBOARD
Study-wise decoding
REWARD
BALLOON
EXPLODE
............
Deep linear model
Joint training
Bayesian deep learning
Resting-state
functional loadings
512 resting-state
networks
Task functional
loadings
128 task-optimized
networks
Contrast
maps
200,000
voxels
Contrast
name
Probability
vector
FACE BODY
HOUSE TOOL
512 components
[Mensch... 2018]
G Varoquaux 28
52. 3 Deep linear model – multi-task – Bayesian
OpenfMRI
HCP
Camcan
Brainomics
...
...
Information flow
Non-negative
matrix factorization
SUCCESSFUL
GO
HOUSE
HORIZONTAL
CHECKERBOARD
Study-wise decoding
REWARD
BALLOON
EXPLODE
............
Deep linear model
Task functional
loadings
128 task-optimized
networks
[Mensch... 2018]
Intermediate representation:
Task-optimized networks
G Varoquaux 28
53. 3 Deep linear model – multi-task – Bayesian
OpenfMRI
HCP
Camcan
Brainomics
...
...
Information flow
Non-negative
matrix factorization
SUCCESSFUL
GO
HOUSE
HORIZONTAL
CHECKERBOARD
Study-wise decoding
REWARD
BALLOON
EXPLODE
............
Deep linear model
Task functional
loadings
128 task-optimized
networks
[Mensch... 2018]
Intermediate representation:
Task-optimized networks
Improves decoding in each study
0% 20% 40% 60% 80%
Decoding accuracy on test set
Stop-signal w/ spoken & manual resp.
The Human Connectome Project
UCLA LA5C
Simon task
Plain or mirror-reversed text
Stop-signal
CamCan audio-visual
Localizer
Cross-language repetition priming
Twin localizer
High-level math
Spatio-temporal judgement (retake)
Constit. struct. of sent. & music
Sentence/music complexity
Word & object processing
Emotion regulation
Arithmetic & saccades
Brainomics localizer
Incidental encoding
Spatio-temporal judgement
False belief
Visual object recognition
Face recognition
Stop-signal & classification
Classification learning
Compression
Motor task & word/verb generation
BART, stop-signal, emotion
Rhyme judgment
Auditory & Visual Oddball
Weather prediction
Stop-signal & classification (retake)
Balloon Analog Risk-taking
Mixed-gambles
Classif. learning & reversal
Task fMRI study
Decoding from
voxels
Decoding from
multi-study
networks
Chance level
-5% 0% 5%10%15%
Multi-study acc. gain
G Varoquaux 28
54. 3 Deep linear model – multi-task – Bayesian
OpenfMRI
HCP
Camcan
Brainomics
...
...
Information flow
Non-negative
matrix factorization
SUCCESSFUL
GO
HOUSE
HORIZONTAL
CHECKERBOARD
Study-wise decoding
REWARD
BALLOON
EXPLODE
............
Deep linear model
Task functional
loadings
128 task-optimized
networks
[Mensch... 2018]
Intermediate representation:
Task-optimized networks
Improves decoding in each study
4 16 32 100 400
Number of train subjects
0%
5%
10%
15%
20%
Multistudyacc.gain
Study
Small studies benefit from large ones
G Varoquaux 28
55. 3 Deep linear model – multi-task – Bayesian
OpenfMRI
HCP
Camcan
Brainomics
...
...
Information flow
Non-negative
matrix factorization
SUCCESSFUL
GO
HOUSE
HORIZONTAL
CHECKERBOARD
Study-wise decoding
REWARD
BALLOON
EXPLODE
............
Deep linear model
Task functional
loadings
128 task-optimized
networks
[Mensch... 2018]
Intermediate representation:
Task-optimized networks
Enables smaller sample sizes
7 15 23 31 39
# training subjects in target study
70%
75%
80%
85%
Testaccuracy
Pinel et al. '07
9 18 28 37 47
70%
80%
90%
Papadopoulos-
Orfanos '12
60 12
60%
65%
C
(Sh
37 47
oulos-
'12
60 121 181 242 302
60%
65%
CamCan AV
(Shafto et al. '14)
Standard decoding Multi-study decoder
2 4 7 9 12
20%
40%
60%
Cohen '09
G Varoquaux 28
56. 3 Deep linear model – multi-task – Bayesian
OpenfMRI
HCP
Camcan
Brainomics
...
...
Information flow
Non-negative
matrix factorization
SUCCESSFUL
GO
HOUSE
HORIZONTAL
CHECKERBOARD
Study-wise decoding
REWARD
BALLOON
EXPLODE
............
Deep linear model
Task functional
loadings
128 task-optimized
networks
[Mensch... 2018]
Intermediate representation:
Task-optimized networks
G Varoquaux 28
57. 3 Deep linear model – multi-task – Bayesian
OpenfMRI
HCP
Camcan
Brainomics
...
...
Information flow
Non-negative
matrix factorization
SUCCESSFUL
GO
HOUSE
HORIZONTAL
CHECKERBOARD
Study-wise decoding
REWARD
BALLOON
EXPLODE
............
Deep linear model
Task functional
loadings
128 task-optimized
networks
[Mensch... 2018]
Intermediate representation:
Task-optimized networks
G Varoquaux 28
58. 3 Deep linear model – multi-task – Bayesian
OpenfMRI
HCP
Camcan
Brainomics
...
...
Information flow
Non-negative
matrix factorization
SUCCESSFUL
GO
HOUSE
HORIZONTAL
CHECKERBOARD
Study-wise decoding
REWARD
BALLOON
EXPLODE
............
Deep linear model
Task functional
loadings
128 task-optimized
networks
[Mensch... 2018]
Intermediate representation:
Task-optimized networks
G Varoquaux 28
59. [Mensch... 2017]
Universal cognitive representations
Multi-task decoding:
- decoding the original task labels
Deep linear models:
- intermediate representations across studies
- statistical power from big to small studies
Face Shape
Reading Shape
Language Music
Saccades Calculation
G Varoquaux 29
60. [Mensch... 2017]
Universal cognitive representations
Multi-task decoding:
- decoding the original task labels
Deep linear models:
- intermediate representations across studies
- statistical power from big to small studies
Face Shape
Reading Shape
Language Music
Saccades Calculation
G Varoquaux 29
62. 4 Beyond heterogeneity: predicting autism across sites
Accuracy
Fraction of subjects used
More data is better (up to 1000 subjects)
[Abraham... 2017]
G Varoquaux 31
63. 4 Brain aging: a surrogate biomarker
[Liem... 2017]
Predicting brain aging = chronological age
Multi-modal: brain connectivity and morphology
Age prediction: mean absolute error of 4.3 years
G Varoquaux 32
64. 4 Brain aging: a surrogate biomarker
[Liem... 2017]
Predicting brain aging = chronological age
Multi-modal: brain connectivity and morphology
Age prediction: mean absolute error of 4.3 years
Discrepency with chronological age
correlates with cognitive impairment
0 2 4
Brain aging discrepancy
(years)
-0.38
0.74
1.72
Objective Cognitive
Impairment group
Normal
Mild
Major
G Varoquaux 32
66. @GaelVaroquaux
Psychoinformatics with machine learning
Prediction for broader theories
AI to model stimuli / the world
Explicit generalization across paradigms
Beyond oppositions
Encoding complete descriptions of tasks
Decoding multiple facets of cognitions
67. @GaelVaroquaux
Psychoinformatics with machine learning
Prediction for broader theories
AI to model stimuli / the world
Explicit generalization across paradigms
Beyond oppositions
Encoding complete descriptions of tasks
Decoding multiple facets of cognitions
Useful with imperfect labels
Extracting common representations
Surrogate biomarkers
68. @GaelVaroquaux
Psychoinformatics with machine learning
Prediction for broader theories
AI to model stimuli / the world
Explicit generalization across paradigms
Beyond oppositions
Encoding complete descriptions of tasks
Decoding multiple facets of cognitions
Useful with imperfect labels
Extracting common representations
Surrogate biomarkers
Software: nilearn
http://nilearn.github.io ni
69. References I
A. Abraham, M. P. Milham, A. Di Martino, R. C. Craddock,
D. Samaras, B. Thirion, and G. Varoquaux. Deriving
reproducible biomarkers from multi-site resting-state data: An
autism-based example. NeuroImage, 147:736–745, 2017.
Y. Bengio. Learning deep architectures for ai. Foundations and
trends in Machine Learning, 2:1–127, 2009.
M. Eickenberg, A. Gramfort, G. Varoquaux, and B. Thirion. Seeing
it all: Convolutional network layers map the function of the
human visual system. NeuroImage, 152:184–194, 2017.
J. V. Haxby, I. M. Gobbini, M. L. Furey, ... Distributed and
overlapping representations of faces and objects in ventral
temporal cortex. Science, 293:2425, 2001.
D. H. Hubel and T. N. Wiesel. Receptive fields, binocular
interaction and functional architecture in the cat’s visual cortex.
The Journal of physiology, 160:106, 1962.
70. References II
A. G. Huth, S. Nishimoto, A. T. Vu, and J. L. Gallant. A
continuous semantic space describes the representation of
thousands of object and action categories across the human
brain. Neuron, 76:1210, 2012.
K. N. Kay, T. Naselaris, R. J. Prenger, and J. L. Gallant.
Identifying natural images from human brain activity. Nature,
452:352, 2008.
F. Liem, G. Varoquaux, J. Kynast, F. Beyer, S. K. Masouleh, J. M.
Huntenburg, L. Lampe, M. Rahim, A. Abraham, R. C.
Craddock, ... Predicting brain-age from multimodal imaging
data captures cognitive impairment. NeuroImage, 2017.
N. K. Logothetis, J. Pauls, and T. Poggio. Shape representation in
the inferior temporal cortex of monkeys. Current Biology, 5:552,
1995.
71. References III
A. Mensch, J. Mairal, B. Thirion, and G. Varoquaux. Learning
neural representations of human cognition across many fMRI
studies. In NIPS, 2017.
A. Mensch, J. Mairal, B. Thirion, and G. Varoquaux. Extracting
universal representations of cognition across brain-imaging
studies. In prep, 2018.
A. Newell. You can’t play 20 questions with nature and win:
Projective comments on the papers of this symposium. 1973.
B. Olshausen ... Emergence of simple-cell remainsceptive field
properties by learning a sparse code for natural images. Nature,
381:607, 1996.
R. A. Poldrack, Y. O. Halchenko, and S. J. Hanson. Decoding the
large-scale structure of brain function by classifying mental
states across individuals. Psychological Science, 20:1364, 2009.
Y. Schwartz, B. Thirion, and G. Varoquaux. Mapping cognitive
ontologies to and from the brain. In NIPS, 2013.
72. References IV
G. Varoquaux and R. A. Poldrack. Predictive models can overcome
reductionism in cognitive neuroimaging. In rev, 2018.
G. Varoquaux and B. Thirion. How machine learning is shaping
cognitive neuroimaging. GigaScience, 3:28, 2014.
G. Varoquaux, Y. Schwartz, R. A. Poldrack, B. Gauthier,
D. Bzdok, J. Poline, and B. Thirion. Atlases of cognition with
large-scale brain mapping. In rev, 2018.