Model-Based Machine Learning for Real-Time Brain Decoding: Neurofeedback derived from real-time functional magnetic resonance imaging (rtfMRI) is promising for both scientific applications, such as uncovering hidden brain networks that respond to stimulus, and clinical applications, such as helping people cope with brain disorders ranging from addiction to autism. One of the greatest challenges in applying machine learning to real time brain “decoding” is that traditional methods fit per-voxel parameters, leading to large computational problems on relatively small datasets. As such, it is easy to over-fit parameters to noise rather than the desired signals. Bayesian model-based hierarchical topographical factor analysis (HTFA) solves this problem by uncovering low-dimensional representations (latent factors) of brain images, fitting parameters for latent factors (rather than voxels) while removing the false assumption that all voxels are independent. In this talk, we’ll discuss the promise of using this and other model-based machine learning to better understand full-brain activity and functional connectivity. And we’ll show how Intel Labs and its partners are combining neuroscience and computer science expertise to further extend such algorithms for real-time brain decoding.
5. Brain Image Analysis/Decoding
5
• Huge amount of data
• 1 volume per scan period (1~2s)
• 100K ~150K voxels per volume
• 100’s ~ 1000’s scans per experiment
• Need sophisticated preprocessing to denoise
• Thermal and system noise from scanner HW
• Head motion, respiration, heart beat, etc., physiological processes
• Neuronal activity related to non-task-related brain process
• Prone to overfitting – typically number of observations < number of features
6. 6
General Linear Model (GLM)
General linear model
Statistical parametric map (SPM)
Design matrix, Sm
Statistical
inference
Realignment Smoothing
Normalisation
Image time-series
Template
Kernel
Y = ( Σ hm conv Sm) + ε
hm
i = bi . βm
i
Haemodynamic Response Function (HRF)
And its partial derivatives
Preprocessing to denoise
8. 8
Brain networks are complicated and dynamic.
Turk-Browne, N.B. (2013) Functional interactions as big data in the human brain. Science 342, 580-584.
9. 9
Can we have a model that describes local and
global spatial dependencies, as well as dynamic
brain networks?
10. 10
Topographic Factor Analysis (TFA)
Manning JR, Ranganath R, Norman KA, Blei DM (2014) Topographic Factor Analysis: A Bayesian Model for Inferring Brain
Networks from Neural Data. PLoS ONE 9(5): e94914. doi:10.1371/journal.pone.0094914
12. 12
TFA discovers latent factors.
Manning JR, Ranganath R, Norman KA, Blei DM (2014) Topographic Factor Analysis: A Bayesian Model for Inferring Brain
Networks from Neural Data. PLoS ONE 9(5): e94914. doi:10.1371/journal.pone.0094914
13. 13
TFA discovers brain networks.
Manning JR, Ranganath R, Norman KA, Blei DM (2014) Topographic Factor Analysis: A Bayesian Model for Inferring Brain
Networks from Neural Data. PLoS ONE 9(5): e94914. doi:10.1371/journal.pone.0094914
14. 14
How can we discover factors common amongst
humans while preserving key individual
differences?
15. 15
Hierarchical Topographic Factor Analysis (HTFA)
Manning JR, Stachenfeld K, Ranganath R, Turk-Browne N, Norman KA, Blei DM. A probabilistic approach to full-brain
functional connectivity. Submitted to PNAS.
16. 16
Graphical Model for HTFA
Manning JR, Stachenfeld K, Ranganath R, Turk-Browne N, Norman KA, Blei DM. A probabilistic approach to full-brain
functional connectivity. Submitted to PNAS.
subject
trials
V voxels
y observed voxel activations
latent factors (µ, )
weights
Individual difference
Global
Factors
17. 17
HTFA Inference Algorithm
while global template not converged and nIter < maxOuterIter do
for subject = 1 to do
while individual factors not converged and mIter < maxInnerIter do
Estimate new weight matrix based on existing centers/widths
Estimate new centers/widths based on existing weights
mIter ++
end
Update global template based on subject’s new centers/widths
end
nIter ++
end
for subject = 1 to do
Update weight matrix based on converged global template
end
18. 18
In essence, TFA/HTFA is a type of factor
analysis. How does it compare with other factor
analyses?
19. 19
TFA/HTFA vs PCA vs ICA
• Commonality
• All decompose observed brain images into a weighted sum of
components
• Difference
• PCA & ICA emphasize the orthogonality or independence of
components. They cannot capture dynamic brain networks
• TFA/HTFA relax the orthogonality/independency requirement, and
with a closed-form factor function, are able to discover richer
information from brain images
• local dependencies
• global dependencies
• dynamic brain networks
22. 22
Bringing HTFA to Reality
Two initiatives:
Reduce the reconstruction error on small number of
factors (K<10) to be lower than 5%
Reduce the overall execution time of a key case study (10
subjects, 10 sources, 200images/subject) to be less than
5mins
23. 23
HTFA reconstruction error was …
Need more optimization when
the number of factors is small
Results are pretty good when
the number of factors is large
24. 24
HTFA reconstruction error is smaller.
Global Centers
Before Optimization
Global Centers
After Optimization
global centers (x) global centers (y) global centers (x) global centers (y)
26. 26
Methods for Speeding up HTFA
Used Intel Math Kernel Library (MKL) where appropriate, e.g.,
single/double precision nonlinear least square solver with/without
constraints
Used thread-level parallelism
Optimized matrix operation order to better utilize cache locality
27. 27
HTFA Speedup Results
0
0.2
0.4
0.6
0.8
1
1.2
1 2 3
Normalized
ExecutionTIme
Raw Data (#factors, #subjects, #img/subject)
HTFA optimization and speedup
Before Optimization
After Optimization
3X to 10X speedup after optimization
28. 28
Recap
Real-time brain decoding can save lives!
Bayesian model-based HTFA is promising
for decoding real-time fMRI data
Intel is working with Princeton to bring real-
time full-brain decoding closer to reality