The submission deadline has been extended to *December 15* for the joint JAES/DAFx Special Issue, "New Frontiers in Digital Audio Effects"! aes2.org/wp-content/u...

Halloween logicals, still the best 10/31 venn

In astronomy, an analogous effect is the Malmquist bias. "Usually heard in the phrase, 'You fool, you forgot to account for the Malmquist bias.'" -- Ethan Vishniac en.wikipedia.org/wiki/Malmqui...

Wasn't Universe 25 the one where all the mice died? @github.com Education, are you saying we're going to eat each other?

Leaderboard Contest! In my Deep Learning class we ran an internal contest for the best latent flow matching generative model for MNIST digits, We're now opening that up to others to submit via GitHub! Details and instructions are on the page: 2025-dlaie-leaderboard-open.streamlit.app

Been working on a tutorial on Guidance for (latent) flow-based generative models, and happened upon a great metaphor today while taking a little "me" time. :-)

Claude has a "weekly limit" now? I'm already a Pro subscriber! :-/

A speech about what drives me, how science and open source are bitter victories, unable to make improve the world if society does not embrace them for the better: gael-varoquaux.info/personnal/a-...

Running a Leaderboard contest in my Deep Learning & AI Ethics class: combine previous lessons into a full latent flow-matching generative model (for MNIST digits). Deadline's tomorrow for students, will later release materials so others can try. 2025-dlaie-leaderboard.streamlit.app

😮🤯 Stoked to see what Anika Nilles & Rush do together! 🤘

Here's (part of) an answer for Exercise 5: namely a fully convolutional VAE with residual connections, for which the latent space is a "wee image": gist.github.com/drscotthawle...

New lesson! "An Arsenal of AutoEncoders (AE, VAE, Conditional VAE, VQ-VAE)" Colab link: github.com/drscotthawle...

I wonder if all the job postings requiring "publication in NeurIPS, ICLR, ICML, or similar conferences" might have anything to do with the massive glut of submissions to these venues.

And lastly, if @neuripsconf.bsky.social would choose to reverse the decisions on the papers affected by space constraints, we would be happy and able to accommodate their presentation

Anyone who wrote a negative review of the new Spinal Tap movie did not get the first one either.

PREACH IT CLAUDE!

The AES AIMLA 2025 conference will feature 4 keynote speakers from field leaders covering topics on ethical audio research, current trends in AI for music, fair generative AI, and commercial AI for the creative industries - see details at: aes2.org/event-extra/...

Today I realized I can ask an LLM to generate Jupyter notebooks (w/ text & code) by having it generate Markdown with cell delimiters and then convert the file format via `jupytext`. (Apparently Cursor does this.) Prompt: "Write a lesson comparing MLP vs CNN for MNIST classification..."

Exercise for the inquisitive: "Let's switch roles. Pretend I'm the LLM and you're the user. Send me prompts and I'll try to fulfill them."

A really nice description (with code examples) of using @openacousticdevices.info AudioMoth and Raspberry Pi together for soundscape monitoring. towardsdatascience.com/audio-spectr...

I often find I can make no progress research-wise until I really, really dumb something down for myself. Not sure if it's a curse or a blessing, but it does help when (later) teaching the topics.

I'm now adding "wacky" hidden-text prompt injections into public-facing documents, syllabi, & more. The goal is not to break or "game" automated reviewing systems, but to help differentiate reviews performed solely by LLM from those that included human oversight.

"Have you tried logging out and logging back in?" has replaced "Have you tried turning it off & on again?" #saas #microsoftoffice

I learned of Afremov during the summer-2022 heyday of Stable Diffusion when freelance/hobbyist researchers shared lists of artist-prompt-style pairings. (Sadly most of the researchers' names got removed due to misdirected backlash) Cool opportunity to support artists' fam afremov.com/deal-of-the-...

"You Are Absolutely Correct! Yes-Men ARE Worse Than Useless as Advisors!" ...is my proposed talk for an upcoming Baylor symposium on "Technology and the Human Person in the Age of AI" (ifl.web.baylor.edu/conferences-...) No apologies. ;-)

Scientists & Academics: "I suppose you COULD use a gigantic LLM for that, but you don't have to. You can train your own much smaller model." Business Folks: "I suppose you COULD train your own small model, but you don't have to. You can just call a gigantic LLM via an API."

My thought when I see people giving LLM agents control over their local machines: www.youtube.com/embed/uj0mtx...

Mission accomplished.

Look, speaking as a physicist, it is hard for me to put into words how batshit crazy this is.

Endorsing: Chrome extension for Claude, to upload entire directory trees to Projects, preserving structure via lengthened filenames. It just works! chromewebstore.google.com/detail/claud...

Exchanged my last $64 USD and got 45 Euros for it. Thanks Donald. (Sure there's a fee but still)

Raymond Fan is a smart guy with great insights into generative models. And, I must add, a fellow Physics PhD. ;-)

Travel in an era when (humans at) information kiosks have disappeared, replaced by too-few ticket vending machines and QR codes to similarly-automated websites: "BUT I'M CONFUSED AND I WANT TO ASK A QUESTION!" All you can do is grab a bystander.🤷‍♂️ @trenitalia.bsky.social

Physics conferences have the "General Theory" session where they send the crackpots. What's the equivalent for ML/NN/CS conferences?

Hi from Rome. 👋 #IJCNN

"increasingly bizarre delusions" may describe my research program! 🤪 The LLMs assure me I'm innovating😉 "Seem to have gotten sucked into dizzying rabbit holes in which the AI acts as an always-on cheerleader and brainstorming partner for increasingly bizarre delusions." futurism.com/chatgpt-ment...

Monday afternoon at #IJCNN I'll give a tutorial on the full dev pipeline for conditional latent flow-matching generative models --we'll talk a LOT about autoencoders! -- plus "inference tricks" e.g. inpainting. This follows Jesse Lai's (Sony AI)'s tutorial on diffusion models. Come add discussion!

Guesses why vector_quantize_pytorch's ResidualVQ might start throwing NaN's late in training, right when my autoencoder is crossing 95% reconstruction accuracy? Maybe something in the orthogonal regularization / rotation trick? @lucidrains.bsky.social @cfifty.bsky.social wandb.ai/drscotthawle...

"You're absolutely right! I made a mistake." --Claude, several times a day. Using LLMs as research assistants assumes a high degree of domain knowledge on the user's part: You need to be able to notice when the model is spouting BS, which is often. A human assistant would likely be less confident.

🤯 This video is 💯! I will probably be telling people at a future lecture to just watch this video instead of listening to me. lol

Further interesting progress from the flow-matching experts in France👇

I think we should be building systems that complement people; systems that do well the things that people do poorly; systems that make individuals and organizations more effective and more humane. 3/

👇

Been visualizing the heck out of my latent flow-matching models lately. Re. "What does Vector Quantization do?": If you just look at the locations of vectors, they may look like a gaussian-shaped "blob". But within that blob can be very different probabilities, vs. the smooth shape of a gaussian.

Call for Papers! AES+DAFx Special Issue, "New Frontiers in Audio Effects" - note that despite the title the list of topics is quite broad & even includes MIR. Deadline: December 1, 2025. Planned publication: Summer 2026. Link to CFP: users.spa.aalto.fi/vpv/JAES_V73...

"Introduction to Mean Flow" - Blog Post + Colab! New tutorial by Raymond Fan (U. Toronto physics PhD). Raymond very clearly illustrates the concepts of Mean Flow and the models' relation to Flow Matching / Rectified Flow. rfangit.github.io/blog/2025/in...

Veo 3: This new horror is "a cat music producer wearing headphones in a tastefully-lit recording studio, singing and bopping its head along to a new pop song 'meow meow meow'."

If you do in fact want Covid boosters for yourself or anyone, you can submit public comment to the FDA here: www.regulations.gov/commenton/FD...

New from U. Edinburgh Acoustics & Audio Group! "Despite their ability to yield perceptually convincing [guitar distortion] models, [neural networks] are susceptible to frequency aliasing ...Nonlinear activation functions create both the desired harmonic distortion and unwanted aliasing distortion."

I wanted to highlight PBDL's brand-new sections on diffusion models with code and derivations! Great work by Benjamin Holzschuh, with neat Jupyter notebooks 👍 All the way from normalizing flow basics over score matching to denoising & flow matching. E.g., colab.research.google.com/github/tum-p...