Max Wasserman

CS PhD at UofR. ex-Apple Applied Scientist intern. BSE at UPenn. Bayesian ML, Optimization. Graph Learning.

[Max Wasserman @MaxWasserman1]

@sp_monte_carlo This aligns with the observation that pointwise evaluation of the posterior is of limited use.

[Max Wasserman @MaxWasserman1]

In most introductory augments, posterior inference is deemed intractable in non-toy settings due to the integral involved in the marginal likelihood p(x) = \int p(x,z) dz.

But even with p(x) in hand, we can only evaluate the posterior p(z|x) = p(x, z) / p(x) pointwise.

[Max Wasserman @MaxWasserman1]

@sp_monte_carlo Answering queries about unobserved variables requires integration.

Integration of a general probability measure is harder than local approximation (differentiation) because it involves the entire space, which in general has uncountable support.

[Accepted papers at TMLR @TmlrPub]

Graph Structure Learning with Interpretable Bayesian Neural Networks

Max Wasserman, Gonzalo Mateos.

Action editor: Fabio Stella.

https://t.co/0IMZg6bm0u

#graphs #graph #deep

[Max Wasserman @MaxWasserman1]

@betanalpha @sp_monte_carlo @rlmcelreath @roydanroy @sirbayes @DavidDuvenaud @tcarpenter216 @bayesian_stats @SolomonKurz These links make me want to go into a cave - armed with a pen, paper, and Papa Rudin - and only emerge when I can extracted every ounce of wisdom from your writings.

[Max Wasserman @MaxWasserman1]

@sp_monte_carlo I have a feeling I'll be chewing on this response for a while. Thanks!

[Max Wasserman @MaxWasserman1]

Bayesian folks, what am I missing? @sp_monte_carlo @rlmcelreath @roydanroy @sirbayes @DavidDuvenaud @tcarpenter216 @betanalpha @bayesian_stats @SolomonKurz

[Max Wasserman @MaxWasserman1]

From what I understand, successful inference typically means “we can now approximate expectations of functions of these unobserved variables z”.

This requires posterior samples. So p(x) is not really the obstacle to successful inference here.

[Max Wasserman @MaxWasserman1]

@StatMLPapers I'll be presenting this work in NYC at the Annual ML Symposium on Oct. 18th. I'll be around for a few days before/afterward—feel free to reach out if you're in town and want to chat!

https://t.co/HchnceuXVJ

@NYASciences @gonchimateos @npparikh @dshieble @sirbayes @phanein

[Max Wasserman @MaxWasserman1]

@StatMLPapers Proud of this work.

The Kumaraswamy has a key role to play in scalable latent variable modeling, but use has been limited by unstable implementations.

We stabilize it and introduce the Variational Bandit Encoder—a simple, scalable approach to contextual multi-armed bandits.

[Stat.ML Papers @StatMLPapers]

Stabilizing the Kumaraswamy Distribution https://t.co/zepvr5EXTg

[Max Wasserman @MaxWasserman1]

@npparikh I now think proper ordering (e.g. baby Rudin before Munkres), periodization (push one area hard at a time, maintain knowledge in others), and a supportive social context (being around people who value such endeavors) are key. Properly executing all 3 is hard.

[Max Wasserman @MaxWasserman1]

@sp_monte_carlo Hot off the presses.

https://t.co/ybxv9C2hio

[Max Wasserman @MaxWasserman1]

@sp_monte_carlo Check out "Model-Based Deep Learning" https://t.co/sy4N5sovwQ

If pdf is needed, dm me.

I have a cool paper coming out on this topic. I link unrolling params to properties of the data, allowing for (interpretable) Bayesian modeling over them. You’ll enjoy it. ~2 weeks to arxiv.

[Max Wasserman @MaxWasserman1]

@Maxematician @sp_monte_carlo For what it’s worth, this is how’s its introduced in Baby Rudin.

[Max Wasserman @MaxWasserman1]

@prof_g I never took your class. You met with me when I reached out to ask silly questions on learning math when I didn’t know up from down. Your kindness, approachability & eccentricity acted as motivation to give the ‘hard’ stuff a shot. And I’m still going!

[Max Wasserman @MaxWasserman1]

@alfcnz @RPstrength SS will take you on a 2 year tangent of ‘hyppe drives’, ‘sets of fahves’, and drinking gallons of milk.

https://t.co/l9wT2HcfI0

[Max Wasserman @MaxWasserman1]

@alfcnz @RPstrength SS stuff will focus on athletics, not changes in body composition (the, err, pear shape you’ve referred to).

SS isn’t bad, but doesn’t compare to RP for your goals.

RP does a good job making clear what does & doesn’t matters in a non-ideological, minimally overwhelming way.