#neurips2023 — Public Fediverse posts

It seems a fair AI would treat you the same if you had a different race/gender/disability/etc., but how can we ever test counterfactual fairness? In #NeurIPS2023 w Victor Veitch, we show you sometimes can with simple, observed metrics like group parity! 🧵 https://arxiv.org/abs/2310.19691

I was an invited speaker at the Neurips conference in New Orleans in Dec 2023 for the NeuroAI social.

I was more than surprised to be invited to what is now primarily an AI/ML conference (despite "Neural" being the first word, and the conference's origins in comp neuroscience). To say that the successful AI systems currently deployed and neuroscience/study of biological intelligence have diverged would be an understatement, it was a somewhat odd choice for the organizers to invite a neurophysiologist like me.

So, I took the invite as an opportunity to talk about attention in biological vision and how whatever they now call as attention in AI/ML/CNN/transformers
is almost orthogonal to what many others and I study within visual neuroscience or psychology or cognitive science.

While the talk was a partial critique of current AI models, it was more a call for them to take seriously the one instance of intelligence (i.e., the biological world) seriously and how it still has much to offer towards designing better AI systems.

If attention is not one of the cognitive ingredients that makes up the intelligence recipe towards autonomous systems, I don't know what is.

The talk slides can be found here: https://www.dropbox.com/scl/fi/927f50bfvqpwtserizgl5/NeuroAI_Neurips_KS2023.pdf?rlkey=r3pgvsyoudwczapjijx80pj7l

#Neurips2023 #NeuroAI #Attention #Vision #BiologicalVision #ActiveVision #SpaceVariance #NonlinearCompression #EyeMovements #Neurodynamics #AutonomousSystems #AI #ML

From #NeurIPS2023: EEG to speech generation. 40% of accuracy in the best case, but if you consider it is a wearable and not fMRI.... "DeWave: Discrete #EEG Waves Encoding for #Brain Dynamics to Text Translation"
https://arxiv.org/abs/2309.14030

#LLM #neuroscience

Research in mechanistic interpretability and neuroscience often relies on interpreting internal representations to understand systems, or manipulating representations to improve models. I gave a talk at the UniReps workshop at NeurIPS on a few challenges for this area, summary thread: 1/12
#ai #ml #neuroscience #computationalneuroscience #interpretability #NeuralRepresentations #neurips2023

Ok, a confession about attending #NeurIPS2023. I was there for the cutting-edge AI/ML innovation and science, sure. But I was *also* there for the food, and to see old friends. But *also*, really maybe the first thing I thought of?

Jazz.

Saw a fantastic concert at Preservation Hall, lots of great music on Frenchman St. And I went to the Jazz museum, really nice.

What I wasn't expecting? Jazz museum is in the old Mint. Which had this very cool old calculator. #SeeImStillNerdy

HuggingFace talking now about Regulatable ML at #NeurIPS2023

My top three voice recognition errors from the #neurips live transcript (ML in structural biology workshop):

3. Kagglers => Cavaliers
2. AlphaFold => Alcohol
1. Generative models => Genital models

#neurips2023

Max Welling: “I’ve been long, long in denial that text based models could help you solve some physics problem… still actually believe that, but never mind. My uncertainty is getting bigger on that one” #neurips2023 #neurips

Outstanding papers at the #NeurIPS2023 Regulatable AI workshop. Congrats to the authors!

Chris Ré in #NeurIPS invited talk: "It's amazing what OpenAI has done. Ilya should get the Turing Award. Maybe not Employee Of The Month. Sorry, couldn't help myself" 😆 #NeurIPS2023

Cutting- edge research
TIES-Merging removes model merging interference #NeurIPS2023 👇

https://arxiv.org/abs/2306.01708

Thanks for the pic @alfcnz
#ML #machinelearning #NLP #nlproc

Oh my Gaaawd #neurips2023 people! How have we been eating fast food when this exists? wow. At Le Bayou on Bourbon Street.

wow

Models can't decide what to do with the update?
Let them vote.
Delete small (noise) parameters
Delete weights with sign disagreements
Sounds interesting?
What if we say model merging\fusion?
https://arxiv.org/abs/2306.01708

Come talk in the poster
#NeurIPS2023
Poster #1118 enjoy the walk...

Check out this #NeurIPS2023 paper by Dinc et al. (2023) who introduce #CORNN, convex #optimization of recurrent neural networks for rapid inference of #NeuralDynamics:

🌍 https://arxiv.org/abs/2311.10200

#CompNeuro #Neuroscience #RNN

In the latest FCAI newsletter:
⭐️ Annual greeting + plans for 2024
🤖 #NeurIPS2023 papers
📢 We are #hiring: PhD + postdoc
📈 ELLIS Institute progress
🟣 Plus more news!
Read here: http://eepurl.com/iFJQuw

cc @ELLISforEurope
#machinelearning #artificialintelligence #phdPosition #ResearchJobs

Learning better with Dale’s Law: A Spectral Perspective - #NeurIPS2023 contribution by Li et al. (2023). It discusses how to train brainlike #RNNs with separate inhibitory and excitatory units with similar performance as standard RNNs:

🌍 https://openreview.net/forum?id=rDiMgZulwi

#RNN #DalesLaw #CompNeuro #Neuroscience

six posters out of 600, is that a good ratio? #neurips2023

This talk was actually one of my #neurips2023 highlights so far, explaining a recently discovered and quite interesting phenomenon that is relatively accessible to non specialists https://en.wikipedia.org/wiki/Double_descent #neurips

Zhongjie presenting our work on tractable circuits for heterogeneous domain called „characteristic circuits“ at #neurips2023.