Today we met the legendary @docmilanfar and in the back @yusuf_dalva looked terrified by his sudden appearance :)

it’s more like CES than any conference I’ve ever attended

"I wasn’t the fastest guy in the world. I wouldn’t have done well in an Olympiad or a math contest. But I like to ponder. And pondering things, just sort of thinking about it and thinking about it, turns out to be a pretty good approach." - Jim Simons

Going from 1k → 2k+ feels like putting on glasses🍌

Come see our poster and chat with @YuyangHu_666, @2ptmvd, and myself. Thursday, poster #4303

Hey #NeurIPS2025 people - don’t spend all your time at the conference. Get out and explore. This is one of the most beautiful parts of California.

automatic transcription of my talk changes "denoiser" to "teen oyster" fine. changing the title of my talk: "teen oysters as building blocks of machine learning" it is

This is a very nice mini-thread:

https://t.co/pcMZeUMqDE

when you see me approaching your poster #NeurIPS2025

Here’s a 3x3 version of what this sub-pixel shifting convolution looks like. I write all this because a lot of these details are often ignored in ML frameworks. This can result in problems when training and using image-to-image deep models for imaging/low level vision tasks 4/4

Since (δx,δy) are small we use a Taylor series: f(x+δx,y+δy) = f(x,y) + δx ∂f/∂x + δy ∂f/∂y Or f(x+δx,y+δy) = [I + δx Dx + δy Dy] f(x,y) That is, a linear filter [I + δx Dx + δy Dy] applied to f(x,y) gets the job done. Dx and Dy are gradient filters in x,y direction 3/4

There’s a nice way to move an image by an arbitrary sub-pixel amount that’s based on a classic idea from optical flow. Suppose we start with an image f(x,y) and want to move by a small amount (δx,δy) f(x,y) → f(x+δx,y+δy) 2/4

How do you move an image by a tiny (sub-pixel) amount? This is an important question because things like resizing and convolution can accidentally do this. To understand this phenomenon and control it, it’s useful to know how to do it on purpose. 1/4

I’ll be at #NeurIPS2025 this week. If you were Reviewer 2 on my paper, please come find me. I just want to talk. I promise.

worth repeating for the uninitiated: Reinforcement Learning is the technical term for "beatings will continue until morale improves"

“It is rare to find learned men who are clean, do not stink and have a sense of humour” - Duchess of Orléans, in praise of Leibniz

it’s weird to see tech reporters posting about attending NeurIPS - surely, they must have something better to do than chase gossip at an ostensibly technical conference

Our massive conferences are increasingly unable to reliably verify findings at scale. This creates a perverse incentive structure that, instead of rewarding truth and accuracy, prioritizes novelty, citation impact, and speed. The body of literature has become like a floating hot

oh well