Batch normalisation appears to be falling out of favour (probably for the best IMO, so many bugs end up being batchnorm bugs😬). One area where it persists is GAN discriminators (e.g. in StyleGAN-T and VQGAN). Are there any other settings where batchnorm is still hard to avoid?

The first text line detection model for historical documents available on Hugging Face : paper+code+models, all open-source huggingface.co/Teklia/doc-ufc… teklia.com/research/publi… #digitalhumanities LITIS

Just published a literature review on scaling laws Epoch AI! I collected a database of scaling laws for different tasks and architectures and reviewed dozens of papers in the scaling law literature. Check out the database! docs.google.com/spreadsheets/d… 🧵

Pete I think this link will be useful: google-research.github.io/seanet/musiclm… the paper: arxiv.org/abs/2301.11325 also they released a dataset: kaggle.com/datasets/googl… it's a shame the actual library is not available but the research article is pretty good

[CV] Learning Good Features to Transfer Across Tasks and Domains P Z Ramirez, A Cardace, L D Luigi, A Tonioni, S Salti, L D Stefano [University of Bologna & Google] (2023) arxiv.org/abs/2301.11310 #MachineLearning #ML #AI #CV [1/2]

Supercontinuum in 100m of 50 micron Fiber youtube.com/watch?v=l6uH0O…

.Talarian's `GPT for Sheets™` is the gift that keeps on giving! 🔥 Look at how easy it is to create personalized content with it, thanks to #GPT3's seamless integration! 🤯👇 Get the add-on here: 🔗workspace.google.com/marketplace/ap…

In the last week, three people on reddit/r/statistics have asked about testing whether a sample came from a Gaussian distribution. The answer is that you should never test for normality. The result is a non-answer to the wrong question. allendowney.com/blog/2023/01/2…

On a Welsh island, a seagull swallows a rabbit whole.

🎉 We are thrilled to announce the release of the latest version of mlforecast a #Python library for Scalable #machinelearning 🤖 for #timeseries #forecasting 🚀 This version comes with exciting new features that are sure to make forecasting even more efficient and accurate 🧵

Cyborgism Research Agenda -Does not try to make GPT an agent -Leverages GPT as a general simulator -LLMs reason from scratch about any situation -Advanced prompting interfaces -Branches GPT chains of thought -Injects variance into human thoughts Post alignmentforum.org/posts/bxt7uCiH…

German AI startup Aleph Alpha claims it will have a 300-billion parameter Large Language Model "Luminous-World" trained this year for “highly complex and critical applications” (GPT-3 has 175 billion parameters)

"MarioGPT: Open-Ended Text2Level Generation through Large Language Models" This looks like a pretty fun and creative project! The best part, it's based on a distilled GPT-3 model and can be trained on a single GPU.

In our paper arxiv.org/abs/2302.10360 on the arXiv today, we present results from a study investigating the energy-efficiency advantage that could be achieved in executing state-of-the-art Transformer models on optical hardware.

This quietly slipped out yesterday. Briefly, the UK is traditinally a strong source of clinical trials but patient recruitment has fallen by 44% (!) due to slow & unpredictable set-up of research sites. Independent review into UK clinical trials buff.ly/3lVWGtx

We've officially passed 200K public datasets!!! 👏👏

One exciting feature of our partnership with AWS is that we now have 1000+ GPUs to start training really large models at Hugging Face 🔥🔥🔥! We’ll be working hard to make closed models open, starting with LLMs and friends 🤓 Which closed models would you like to be open?

We're excited to welcome FuseMedML to the PyTorch ecosystem! FuseMedML is part of the BiomedSciAI organization which provides tools for AI-based accelerated discovery of biomarkers and molecules in the biomedical domain. See the latest here: hubs.la/Q01D6TZ70

Robust Hybrid Learning With Expert Augmentation Antoine Wehenkel, Jens Behrmann, Hsiang Hsu, Guillermo Sapiro, Gilles Louppe, Joern-Henrik Jacobsen. Action editor: Jasper Snoek. openreview.net/forum?id=oe4dl… #expert #models #modelling

Recitation-Augmented LLMs -Usually: Retrieve relevant docs & give to LLM to answer -Instead: Recite relevant passages from LLMs' own memory via sampling, then produce final answer -State-of-the-art on closed-book Q&A Paper arxiv.org/abs/2210.01296 Code github.com/Edward-Sun/REC…