The secret sauce works! www.forbes.com/sites/daveal...

Blog post by A-Tang Fan and Doug Watt about Axelera.ai's Voyager SDK: community.axelera.ai/product-upda...

I'm delighted to share that I joined the Axelera team this week to deliver the next generation AI compute platform. axelera.ai

Reinforcement Learning (RL) just landed a stellar breakthrough with reasoning language models. Yet, RL has a distinctly bad reputation. See “To RL or not to RL” (www.reddit.com/r/MachineLe...) on reddit. I'd like to revisit the basic math of RL to see why. Let's enter the dungeon!

Are you still using LoRA to fine-tune your LLM? 2024 has seen an explosion of new parameter-efficient fine tuning technique (PEFT), thanks to clever uses of the singular value decomposition (SVD). Let's dive into the alphabet soup: SVF, SVFT, MiLoRA, PiSSA, LoRA-XS 🤯...

Well worth reading: @fchollet.bsky.social 's analysis of OpenAI's o3 breakthrough score of 76% on the ARC-AGI benchmark: arcprize.org/blog/oai-o3-...

Sakana.ai's Transformer² arxiv.org/abs/2501.06252 paper features a cool new parameter-efficient fine-tuning (PEFT) technique that makes tuned models composable!  Let’s dive in 💦. (They have stunning artwork on their website 🤩 too)

Posted on Twitter in Nov: looking at the "AI achieves Kaggle Grandmaster Level" paper published last week: arxiv.org/abs/2411.03562. A massive 88-page paper. Here is a summary.

I'm exiting Twitter/X after being served an ad there for a neo-nazi podcast. The X exodus is massive and you don't have to lose your followers. Thanks to #HelloQuitX I've registered 19567 new passengers for a journey to #BlueSky. Join us on app.helloquitx.com.

Personal update: I am no longer at Hugging Face. I will take some time to pursue personal projects and find my next adventure. DMs open. Feel free to ping me if you have an interesting AI/ML project to share!

Did you know that you can load the newest checkpoints (like Llama 3.2) into Keras directly from the original HuggingFace release (safetensors)? I tried - and lived to tell the tale: huggingface.co/blog/keras-l...

The fourth nice thing we* have for you this week: PaliGemma 2. It’s also a perfect transition: this v2 was carried a lot more by @andreaspsteiner.bsky.social André and Michael than by us. Crazy new sota tasks! Interesting res vs LLM size study! Better OCR! Less hallucination!

Pitting a few Keras LLMs against each other: huggingface.co/blog/keras-c... Using an super-simplified scenario, I wanted to see how easy it is to get them to fix their own mistakes.

Pitting a few Keras LLMs against each other: https://huggingface.co/blog/keras-chatbot-arena Using an super-simplified scenario, I wanted to see how easy it is to get them to fix their own mistakes.

I'm looking at the "AI achieves Kaggle Grandmaster Level" paper published last week: https://arxiv.org/abs/2411.03562. A massive 88-page paper. Here is a summary.

Looking at the code of the recent moonshine release: https://github.com/usefulsensors/moonshine (a speech recognition model optimized for mobile devices). It has a very clean Keras implementation! A couple of noteworthy details: (1/5)🧵

Did you know that you can load the newest checkpoints (like Llama 3.2) into Keras directly from the original HuggingFace release (safetensors)? I tried - and lived to tell the tale: https://huggingface.co/blog/keras-llama-32

I just noticed: Keras 3 is now the default in Colab. Nice! Keras+JAX, Keras+PyTorch, Keras+TF right at your fingertips.

Announcement #2: new Keras + Hugging Face integration: you can now load HF fine-tuned models through Keras, even if they have not been fine-tuned in Keras. As long as the architecture is implemented in KerasNLP, weights will be converted on the fly. Colab:...

Gemma 2 has landed in KerasNLP: https://developers.googleblog.com/en/fine-tuning-gemma-2-with-keras-hugging-face-update/

New Keras starter notebook from @awsaf49 for the "LMSYS Chatbot Arena Human Preference" competition on Kaggle. This one is interesting for how it achieves preference classification over pairs of (prompt+response) using the DeBERTaV3 model from...

Google I/O is today and "Large language models with Keras" is at 4:30PM 🔥with @smn_sdt and @GabrielRasskin !

Just released: you can now upload your Keras models to Kaggle Models or HuggingFace, directly from the Keras API: https://developers.googleblog.com/en/publish-your-keras-models-on-kaggle-and-hugging-face/

New Kaggle starter notebook from @awsaf49 for the Automated Essay Scoring competition: https://www.kaggle.com/code/awsaf49/aes-2-0-kerasnlp-starter. It showcases the right way to do ordinal regression in Keras, i.e. how to predict ordered integer grades reliably (no it's neither a...

A nice blog post about Keras 3 from NSF Unidata https://www.unidata.ucar.edu/blogs/news/entry/why-is-the-keras-3 The post's conclusion: "For deep learning training I (Thomas) will be using a Keras 3 API exclusively. It more closely resembles the scikit-learn api and I find it to be...

Gemma in Keras: how to build a chatbot and fine-tune it to speak like a pirate 🏴‍☠️🦜. This was a fun demo to make! It runs with Keras on JAX with the new keras.distribute.ModelParallel API. Colab: http://bit.ly/gemma-pirate-demo Video: ...

Not one but two Gemma competitions are currently live on Kaggle. And we have Keras starter notebooks for both: https://www.kaggle.com/code/awsaf49/prompt-recovery-with-gemma-kerasnlp-starter https://www.kaggle.com/code/awsaf49/kaggle-qa-with-gemma-kerasnlp-starter Have fun with...

Starter notebook for Kaggle competition "Learning Agency Lab - PII Data Detection", using Keras 3 and Keras NLP. https://www.kaggle.com/code/awsaf49/pii-data-detection-kerasnlp-starter-notebook This is about flagging PII (Personally Identifiable Information), think names, phone...

Keras Starter Notebook alert: "Hacking the human vasculature in 3D": https://www.kaggle.com/code/awsaf49/sennet-hoa-kerascv-starter-notebook-train

Here is a Keras starter notebook for the "Harmful Brain Activity Classification" competition by the legendary @awsaf49. Enjoy! https://www.kaggle.com/code/awsaf49/hms-hbac-kerascv-starter-notebook

The "Self-Extend" paper http://arxiv.org/abs/2401.01325 promises magic for your LLMs: extending the context window beyond what they were trained on. You can take an LLM trained on 2000 token sequences, feed it 5000 tokens and expect it to work. Thread 🧵 (SWA below=sliding window...

Announcement: all KerasCV and KerasNLP models can now be found in Kaggle Models. With nice UI touches: https://www.kaggle.com/discussions/product-feedback/467024 In competitions, this means you can use them with "internet OFF". Yolov8, Segment Anything, Stable Diffusion and more......

Keras is the ideal framework for sharing pre-trained models because it lets your users programmatically inspect and/or edit the models. Tutorial: https://keras.io/examples/keras_recipes/packaging_keras_models_for_wide_distribution/ No more "if you want to do that, you'll have to...

Good looking Jax coders! Seriously, this is a Keras 3.0 demo, runnig KerasCV and KerasNLP generative models on Jax, TensorFlow or PyTorch. https://colab.research.google.com/drive/1tlQ8xOAkrsO34yGoNYVfRAsXbkKoxurS?usp=sharing It also shows how to install everything.

A new KerasCV starter notebook for the Kaggle competition "UBC Ovarian Cancer Subtype Classification and Outlier Detection (UBC-OCEAN)" https://www.kaggle.com/code/aritrag/kerascv-train-and-infer-on-thumbnails/notebook Fun fact: at actually runs on JAX

Fun Kaggle competition: can you score essays from typing skills alone? This dataset gives you anonymized keyboard events where every keypress is "q" and asks you to grade the essay. Here is a KerasNLP starter...

XLA speedups of up to 40x for some models! Epic benchmarking report of Keras models by @RisingSayak https://wandb.ai/sayakpaul/keras-xla-benchmarks/reports/XLA-Compatibility-of-Vision-Models-in-Keras--Vmlldzo0NzQzNDYx

Keras now trains the same models with the same source code and checkpoints on TensorFlow, PyTorch and JAX. "Keras Core" presentation by @fchollet https://www.youtube.com/live/_5fTPEoeFZk?feature=share

Better with a demo: https://bit.ly/keras-on-jax-demo

Great new Object Detection tutorial by @luke_wood_ml, using KerasCV: https://keras.io/guides/keras_cv/object_detection_keras_cv/ A couple of lines to instantiate an object detection model. It's Keras so all the internals of the model are accessible and KerasCV offers easy...

This is what we have been working on for the past year: https://youtu.be/K2PKZS1fPlY KerasCV and KerasNLP. New v0.5 released yesterday!

Visual Language Models (VLMs) have many advantages but one of them is that they are just better at traditional computer vision tasks. Here is a pic from the original CLIP paper, comparing CLIP and good old ResNet101 for image classification:

DINO v2 (https://arxiv.org/abs/2304.07193) has learned about object parts: heads, wings are recognized across distinct categories such as birds and planes! 100% unsupervised! (context: see my DINO self-supervised explainer here https://x.com/martin_gorner/status/1651883371536973825)

For unsupervised training fans, the DINO paper from 2021 is well worth a read. (They recently updated the approach with DinoV2. See comparison video here: https://ai.facebook.com/blog/dino-v2-computer-vision-self-supervised-learning/)

You can hear the gears clicking into place: https://writings.stephenwolfram.com/2023/01/wolframalpha-as-the-way-to-bring-computational-knowledge-superpowers-to-chatgpt/ 🤯

I'm trying to understand the very impressive GPT-4 example from this paper: https://arxiv.org/pdf/2303.12712.pdf titled "Sparks of Artificial General Intelligence: Early experiments with GPT-4"

A fascinating overview of research into "Dataset Distillation" https://arxiv.org/abs/2301.07014 How to train a neural network on fewer data points and achieve the same performance as when training on the original dataset.

Large Language Models are getting good at formal logic: https://arxiv.org/abs/2212.13894 LAMBADA: Backward Chaining for Automated Reasoning.

How can you probe what a language model knows ? If you ask it directly, it might lie (for example because you prefixed your question with untruths, or many other reasons). Contrast-Consistent Search (CCS) gives a way: https://openreview.net/pdf?id=ETKGuby0hcs

Here is @luke_wood_ml explaining Stable Diffusion at #Devoxx https://youtu.be/KaR51-pwzlU#DevoxxStable Diffusion is the first end-to-end model in the new KerasCV library.