Twitter Beğeni Hilesi

Oldies but goldies: M. Belkin, P. Niyogi, Laplacian Eigenmaps for Dimensionality Reduction and Data Representation, 2003. Non-linear dimensionality reduction by embedding data points using the eigenvectors of a graph Laplacian as coordinates. en.wikipedia.org/wiki/Nonlinear…

LLMs excel at in-context learning; they identify patterns from labeled examples in the prompt and make predictions accordingly. Many believe more in-context examples are better. However, that's not always true if the early ascent phenomenon occurs.

My favorite illustration of how geometric intuition fails miserably in high dimensions.

I have put online the content of 10 lectures of 2h on the basics maths of AI. It contains raw transcripts, so probably not super useful, but it might be helpful to give some ideas. mathematical-tours.github.io/maths-ia-cours…

in addition to coffee and cappuccino, apparently this place in Madison also sells solutions to approximation problems with multilayer perceptrons

I am excited for this upcoming talk by Andrew about "optimally" exploring given some offline data! Bonus: We'll hear about the gap between verifiable and unverifiable learning! I hope to see you tomorrow!

Check out our label efficient SFT paper for examples of the clever prompting methods:x.com/rdnowak/status…

Here’s my take on the “mathematical foundations” of machine learning and AI. These course notes cover the basics of statistical learning theory, optimization, and functional analysis. nowak.ece.wisc.edu/MFML.pdf

Oldies but goldies: Frank Rosenblatt, The Perceptron-a perceiving and recognizing automaton, 1957. The first appearance of a neural network (single layer), later generalized to multiple layers. en.wikipedia.org/wiki/Perceptron

Oldies but goldies: Alfred Haar, Zur Theorie der orthogonalen Funktionensysteme, 1910. The first wavelet transform. Later generalized by Ingrid Daubechies. en.wikipedia.org/wiki/Haar_wave…

LLMs can memorize training data, causing copyright/privacy risks. Goldfish loss is a nifty trick for training an LLM without memorizing training data. I can train a 7B model on the opening of Harry Potter for 100 gradient steps in a row, and the model still doesn't memorize.

Donald Trump’s former Sec. Def. Mark Esper: “[Trump] was suggesting that...we should bring in the troops and shoot the protesters.” Q: “The commander-in-chief was suggesting that the U.S. military shoot protesters?” Esper: “Yes, in the streets of our nation’s capital.”

Thanks for the nomination VentureBeat. It's an honor to be included in this glowing list of outstanding women in AI. UW School of Medicine and Public Health UW-Madison BME UW Radiology Integrated Diagnostics and Analytics (IDiA) Lab

This 👇

I just found this buried in my copy of Infinite Jest.

We are pleased to announce plans for a special issue of Signal Processing Magazine focused on the mathematics of deep learning: signalprocessingsociety.org/blog/ieee-spm-… We look forward to your submissions!

Joe Shenouda led this project. We learned a lot about how multi-output neural nets differ from single-output nets. Interesting consequences for network compression. jmlr.org/papers/v25/23-…

Does training a generative model on its own synthetic data always result in MADness/model collapse?Turns out it doesn’t! We show that a diffusion model can “self-improve” using its own synthetic data while preventing MADness/model collapse altogether! Link to paper:

We are excited to launch the first ever weekly caption contest on Toondeloo. Submit your caption at toondeloo.com over the next week and come back to vote. Please share with others, the more the merrier! Here is this week's cartoon.