Matt Jordan

Making datasets at @allen_ai

[Matt Jordan @rev_bucket]

@percyliang Oops! I accidentally uploaded this test set to dolminos last November (fixed now: https://t.co/NN5iVDUwjU)!

And to be crystal clear -- Olmo2 never saw the test set, this was purely a mistake in uploading data for release!

[Matt Jordan @rev_bucket]

@i_am_sidward @AlexGDimakis @docmilanfar (3/3)
3. Another promising direction is to use variants of randomized smoothing, which convolves the network output with some distribution (e.g. gaussian), for which you can also get lipschitz bounds for free

[Matt Jordan @rev_bucket]

@i_am_sidward @AlexGDimakis @docmilanfar (2/3)
2. More promising is to design a net with stability in mind. There's some work which enforces the output net has Lip(f)<=1. Search "Lipschitz Neural Networks" on GScholar for some recent papers. You get worse performance, but get robustness guarantees for essentially free

[Matt Jordan @rev_bucket]

@TheGregYang @AlexGDimakis (2/2)
The two main approaches that might be helpful are:
1) using loose upper bounds to quickly reject candidate polytopes
2) training the nets to be conducive to efficient verification (see also: https://t.co/PvrfS36io0)

[Matt Jordan @rev_bucket]

@TheGregYang @AlexGDimakis (1/2) Unfortunately, if we want to deterministically compute the exact verification, we can't get around the hardness: so scaling will always be an issue for deep nets.

[Matt Jordan @rev_bucket]

Wrote a paper about a geometric technique for exactly computing pointwise robustness of neural nets in arbitrary convex norms: to the best of my knowledge, this is the first complete/exact technique for computing l2 robustness.

[Alex Dimakis @AlexGDimakis]

New paper: Provable Certificates for Adversarial Examples: Fitting a Ball in the Union of Polytopes. Our geometry result: Polytopes for ReLu nets are `perfectly glued' i.e. partition space into a polyhedral complex.
https://t.co/XSVTP57QRo
Code: https://t.co/cIK7V0Z6dF

[Matt Jordan @rev_bucket]

Wrote a paper. Finally pushed it to the arxiv. Use my toolbox for adversarial examples in pytorch!

[Alex Dimakis @AlexGDimakis]

Our new preprint: Quantifying Perceptual Distortion of Adversarial Examples
https://t.co/zwEnZUSgdZ
which introduces a perceptual threat model and a PyTorch toolbox for adversarial ML. Meet #MisterEd
https://t.co/DPopNcuj39

[Matt Jordan @rev_bucket]

#redshift crashing: A fatal error occurred. The database will be restarted.
And the only info I can get is this: https://t.co/90KzPb7N7n

[Rahul Pathak @rahulpathak]

@mystelynx please email me at [email protected]. happy to help.

[Matt Jordan @rev_bucket]

When you have a cooking pot designated solely for filling with ice and sticking your foot in #justRunnerThings

[Matt Jordan @rev_bucket]

How people said #Yolo in the pre-women's suffrage era: http://t.co/8jJG9I3CPp

[Matt Jordan @rev_bucket]

Me when I complete AWS 2FA right before the timer runs out: http://t.co/tqXrIOLs87

[Matt Jordan @rev_bucket]

No better way to cripple a tech startup on a Monday morning than a weekend wifi password change. #guessingGame #bruteForce #starbuckStartup

[TechCrunch @TechCrunch]

With $1.75M In Funding, MaestroIQ Promises Smart Recommendations For Mobile Marketers http://t.co/1uqzzvCNFH by @anthonyha

[Matt Jordan @rev_bucket]

I can't do any work with classical Wiener Measure without snickering. I guess that's what they mean by 'mathematical maturity'

[Matt Jordan @rev_bucket]

After using detexify to find the symbol used to describe the Riemann ZETA function, I realize I have no business doing math #IAmNotASmartMan

[Matt Jordan @rev_bucket]

Pretty much my first experience lobster diving: https://t.co/Xrku0swcqV

[Matt Jordan @rev_bucket]

T. Swizzle on that deep house remix. All about it. https://t.co/dXEcgKPhQ0 #ShakeItOff

[Matt Jordan @rev_bucket]

The fire alarms in the office keep getting tested. Ruins the point of alarms, but makes for a nice strobe to my coding soundtrack. #strobe

[nihal @nihalmehta]

driving to the hamptons with the rockstar @AdTrib_Team