Leo Duan

University of Florida Associate Professor

[Leo Duan @leo_duan]

@raybai07 That’s very nice. Despite my last jokingly phrased tweet, I sincerely believe AI is something we need to openly work with, not against, in our teaching and research. Thanks for sharing how you navigate through this.

[Leo Duan @leo_duan]

There is a growing debate about the use of AI in our communities, so I'd like to share a few thoughts on the *human aspects* based on my interactions with peer researchers and students:
1. Let's stop pretending we didn't use AI in the making of a paper or product, please, just stop.
2. Our contributions are not reduced to zero by the use of AI. A paper is still a manifestation of our ideas, careful presentation & critically, prudent checks against AI slops.
3. Expect work at AI-level productivity. 10 ideas for simulations? Feed them all to AI and let's see results tomorrow morning.
4. AI likes to flood our screen with inside-the-box information (cynically, to sell more tokens), innovation needs clear, original & independent thinking: that's on us.

[Leo Duan @leo_duan]

@GuanyangW @MoMoJing131 Lean is great, but AIs have also found ways to game the compile system, so Lean check isn't 100% bulletproof: https://t.co/g94e4XXNAn

[Leo Duan @leo_duan]

More on the ILP embedding trick: if the combinatorial set (polyhedron) has all vertices integer-valued, then argmax_z (z' \zeta) will a.s. transform a continuous \zeta (blue) onto an int vertex (red). In 1-dim case with binary z, argmax_z (z\zeta) reduces to sign(\zeta). (3/n)

[Leo Duan @leo_duan]

For modeling binary outcomes (e.g. classification, network edges), Albert & Chib (1993) had a brilliant idea: treat each binary as the sign of a latent Gaussian. But what if each outcome is a binary vector obeying combinatorial constraints? (1/n)

[Leo Duan @leo_duan]

(e.g. skip-logic survey data, observed matchings on a bipartite graph) Our new paper (https://t.co/QrLm4XiAYN) replaces the sign function with an integer linear program (from combinatorial optimization) �� a new class of GLMs suitable for structured combinatorial data! (2/n)

[Leo Duan @leo_duan]

@GuanyangW @nawaf_bou On existing HMC packages, we recently explored numpyro, and found it to be amazing — especially on its implementation of Mixed HMC that allows easy update of discrete parameters

[Leo Duan @leo_duan]

Highlighting our recent theory work: "all models are wrong, but some are consistent" 🙌 -- a potentially misspecified Bayesian clustering model can yield an asymptotically correct posterior estimate, including the number of clusters! https://t.co/puQwstcnsP

[Leo Duan @leo_duan]

@sp_monte_carlo No worries. On the modeling front, I'd say if a Bayesian model is inspired by some optimization approach involving a non-smooth part, then there are multiple ways to assign a distribution around the point estimate -- Gibbs posterior is often not the best. DM me if interested.

[Leo Duan @leo_duan]

@sp_monte_carlo I'd say non-smoothness is a "symptom", not the cause. When it comes to MCMC, non-smooth density can be annoying, although there are ways to "smooth it up", e.g. proximal MCMC.

[Leo Duan @leo_duan]

For those of you who dislike the job of correcting format errors/case inconsistency in bibtex file (reference database for latex), I made a small script that does that for you: https://t.co/iK6oyp0Gf8

[Leo Duan @leo_duan]

@raybai07 I'd say people in our field judge based on the level of energy a person shows -- based on that standard, you're still in the 20s. 🤣

[Leo Duan @leo_duan]

Importantly, we've worked really hard on improving & simplifying the algorithm. Now the simple R code runs super fast, and works well in clustering most data. Try it out on your data, and let us know! https://t.co/2T1T9dwGAK

[Leo Duan @leo_duan]

You might know "Normal mixture <=> Generative model for K-means", and I want to tell you "🌟Forest model <=> Generative model for spectral clustering🌟". Check out my joint work with Arkaprava Roy: https://t.co/sUKDZu4uOs

[Leo Duan @leo_duan]

(2) we can now easily extend a vanilla forest model to a hierarchical model, e.g. imagine a low-dimensional z_i as from a forest model, then this z_I is transformed to high-dimensional face photos y_i. We get 80%-90% clustering accuracy.

[Leo Duan @leo_duan]

This generative model buys us 2 things: (1) it tells us the uncertainty (i.e. the "variance") around the point clustering estimate, e.g. how likely two points are in the same cluster? how many clusters are there? etc.

[Leo Duan @leo_duan]

It was such a wonderful experience to be in-person at @ISBA_events . Having real floor / Mont-top discussions on stats and non-stats things reminds me of how good the academic life can be. Already looking forward to the next event!