SILO Seminar - Nov 2023
Imperceptible perturbations (to humans) to inputs can compromise ML systems
Imperceptible perturbations (to humans) to inputs can compromise ML systems
Estimate the risk of a model from training samples
Imperceptible perturbations (to humans) to inputs can compromise ML systems
Estimate the risk of a model from training samples
How should standard restoration approaches be adapted to exploit deep learning models?
Imperceptible perturbations (to humans) to inputs can compromise ML systems
Estimate the risk of a model from training samples
How should standard restoration approaches be adapted to exploit deep learning models?
PART I
Local Sparse Lipschitzness
PART II
Learned Proximal Networks
PART I
Local Sparse Lipschitzness
Setting
Hypothesis Class
Hypothesis Class
Sensitivity
"Complex Deep Nets are locally simple"
"Complex Deep Nets are locally simple"
Adversarial Robustness
Generalization
Learning to solve Inverse Problems
active
weakly inactive
strongly inactive
active
weakly inactive
strongly inactive
active
weakly inactive
strongly inactive
active
weakly inactive
strongly inactive
Theorem:
active
weakly inactive
strongly inactive
Theorem:
active
weakly inactive
strongly inactive
Theorem:
Lemma:
Theorem:
Margin vs stability
Certified Accuracy
Perturbation Energy
0.0
0.5
1.0
1.5
0.0
0.2
0.4
0.6
0.8
1.0
Muthukumar & S. (2023). Adversarial robustness of sparse local lipschitz predictors. SIAM Journal on Mathematics of Data Science, 5(4), 920-948.
Generalization Gap:
more sparsity, lower sparse norms
Lemma:
Lemma:
Main result (informal)
empirical (margin) risk
sparse loss
deviation from prior
PAC-Bayes Bounds
Observation:
Main result (informal)
Is the empirical margin risk smaller than ?
Sparse local radius large enough (all layers) over the sample?
empirical (margin) risk
sparse loss
deviation from prior
Main result (informal)
empirical (margin) risk
sparse loss
deviation from prior
Related results
Examples
Muthukumar & S. (2023). Sparsity-aware generalization theory for deep neural networks. In The Thirty Sixth Annual Conference on Learning Theory PMLR.
Increasing width
PART I
Local Sparse Lipschitzness
PART II
Learned Proximal Networks
PART II
Learned Proximal Networks
measurements
reconstruction
I'm going to use any neural network
... that computes a proximal, right?
That computes a proximal, right?
This is a MAP denoiser...
Let's plug-in any off-the-shelf denoiser!
Ongie, Gregory, et al. "Deep learning techniques for inverse problems in imaging." IEEE Journal on Selected Areas in Information Theory 1.1 (2020): 39-56.
Proposition
Convergence:
(we don't know !)
How do we train so that ?
(we don't know !)
Theorem (informal)
How do we train so that ?
What's in your prior?
MNIST
Fang, Buchanan & S. (2023). What's in a Prior? Learned Proximal Networks for Inverse Problems. arXiv preprint arXiv:2310.14344.
Ram Muthukumar
JHU
Zhenghan Fang
JHU
Sam Buchanan
TTIC
NSF CCF 2007649
NIH P41EB031771
[ Mairal et al., '12 ]
Theorem:
[ Mehta & Gray, '13; S., Muthukumar, Arora, '20 ]