Decades before the modern scaling laws, this paper showed that neural networks behavior under scale follows remarkably predictable laws.
In 1993, researchers at Bell Labs were grappling with a constraint that feels entirely familiar (and contemporary): datasets were outgrowing the available hardware, and training a model to the end was becoming too expensive. To evaluate an architectural tweak to a state-of-the-art model (at the time it was LeNet) on 60,000 samples meant burning up to three weeks of compute time.
To save compute, people would train candidate architectures on small subsets of the data, assuming that the top performer at small scale would remain the top performer at full scale. But with our future wisdom, we know this is not the case.
In "Learning Curves: Asymptotic Values and Rate of Convergence (NeurIPS 93)", using insights from statistical mechanics, they proposed a practical and principled method for predicting the performance of classifiers trained on large datasets (at the time, models were assumed to be large enough). The method was based on a simple power-law modeling of the expected training and test errors.
It is often noted that many of today's breakthroughs in AI and deep learning are actually decades-old concepts that simply lacked the computational power to be tested at the time. While there is some truth to that, it highlights a more valuable lesson: there is immense worth in revisiting early literature and reflecting on foundational ideas we may have prematurely left behind.
So, go explore and find your own inspiration. The current trend has enough champions already!
It has been more than a decade now since the knowledge distillation paper came out.
Knowledge Distillation (KD) is one of my favorite topics, but I have to confess that I'm not a huge fan of the term because I find it confusing (or at least, it has became so over time).
The idea behind KD is not novel; it was there almost a decade before the paper came out (and arguably even a decade before that, back to 1990-91). But this paper is the one that clicked, the one that made the topic much more popular and introduced it to a broader audience.
First, the timing and the authors played a big role: we have Geoffrey Hinton, Oriol Vinyals, and Jeff Dean here. And second, Geoffrey Hinton is really good at idea branding: Model compression?! No, no, no! Let's call it "Knowledge Distillation" and use evocative terms such as "Dark Knowledge" to describe what is being transferred.
It's a great name, but as time has passed, the term became a bit of a relic. KD is no longer solely about compression (KD used to be introduced as a method for model compression, but now model compression is just one application of KD). And the other thing is that the word "distillation" implies some sort of potency here, that the student is somehow more powerful than the teacher, which is not the case (but many counterarguments could be made, for example, more powerful compared to another model trained with no teacher)
Nevertheless, the paper is incredibly well-written, short, and fun to read. It's one of few papers that I read several times. Check it out, and maybe share your thoughts on the topic with us here!
If you had to choose another name for Knowledge Distillation, what would it be?
What if you could train a model on just 10 images instead of 60,000 and still get close to the same performance?
Traditional machine learning requires thousands, even millions, of data points to achieve high accuracy. But what if we could "distill" the entire dataset into just a few synthetic samples?
This is what Dataset Distillation offers. Unlike traditional knowledge distillation, we keep the model fixed and distill the knowledge contained in a massive training set into a tiny set of synthetic distilled images.
The goal is to train a model on this ultra-small set and achieve performance that almost matches what the same model would get when trained on the massive original dataset.
For example, training on only 10 distilled MNIST images (this is equivalent to a single image per class) yields 94% accuracy, compared to 99% when training on the full 60,000 images.
Interestingly, these distilled images look significantly different (as you can see in the image below) from natural images because they are optimized for model training rather than for matching the correct data distribution.
But that's not all.
Most importantly, this same method opens the door to a potent form of data poisoning. Because distilled images are specifically optimized for rapid learning, an attacker can create a tiny set of adversarial distilled images to cause a well-trained model to forget or misclassify a specific category.
What I find fascinating about dataset distillation is this: it mimics human-like learning by letting a model grasp a concept from a single example, but it does so using alien synthetic images that mean absolutely nothing to a human eye!
Wan2.2-I2V-Fast with highly upscaled sequential frame sampling is now available as a Spaces demo, built using Wan2.2-I2V and FLUX.2-Klein. Try the demo using the links below.π
It was supposed to be a failed experiment. Instead, it led to the discovery of one of the most intriguing phenomena in neural networks, simply because a researcher forgot to turn it off and left it running....for a week!
In 2022, researchers at OpenAI were studying how neural networks generalize from their training data. For this task, they were training small transformer models to perform modular arithmetic.
The thing is, neural networks are weird. When a model has an abundance of parameters (like neural nets), it can easily overfit. It essentially memorizes its training data, scoring a perfect 100% accuracy when tested on it, but remains completely clueless when faced with any new instances not present in the training set (close to 0 accuracy). It is like memorizing 1 + 2 = 3 without understanding the concept of addition, so if 2 + 3 wasn't in the training set, the model fails miserably!
Usually, when a model overfits like this, people just cut their losses, turn off the experiment, and move on with their lives.
But sometimes they forget. And that is exactly what happened to our researchers at OpenAI. A week later, they checked back in, and a miracle had happened!
They discovered Grokking (And no, this has nothing to do with xAI's Grok , the term was originally coined by sci-fi author Robert Heinlein to mean understanding something so deeply that it becomes part of you). Grokking is when a neural network suddenly and abruptly learns to generalize long after it has overfitted. Just take a look at the graph in the image below!
Human brains don't recreate every pixel to understand the world!
Most current models in genomics, proteomics, and single-cell transcriptomics rely on generative objectives like masked language modeling or next token prediction. While effective, these architectures waste significant capacity reconstructing raw, noisy sequence details that may not carry functional biological meaning.
But a promising, more efficient alternative is emerging: Joint-Embedding Predictive Architecture (JEPA)
Originally introduced by Yann LeCun for computer vision, JEPA is a non-generative, self-supervised learning (SSL) framework. Instead of predicting raw inputs, it operates as a world model that predicts abstract semantic embeddings in latent space.
Recently, the JEPA framework (and its more efficient LeJEPA variant) has been adapted into the biological sciences to develop performing foundation models and to improve on already existing ones.
It's interesting how each adaptation modified and tailored JEPA to suit its specific biological domain, whether by experimenting with different backbones or complementing the objective with other loss terms.
For example, JEPA-DNA and ProteinJEPA used JEPA as a continual pre-training framework to enhance existing foundation models without training from scratch, while Cell-JEPA and JEPA-DNA employed a hybrid objective that combines the JEPA loss with a traditional language modeling loss.
The article below provides an overview of these implementations, along with others that came out this year. As always, your thoughts and feedback are welcome and highly appreciated!
Released lafzyn , built over Qwen, an Urdu language model that converts Urdu text into IPA phonetic transcription, with GGUF builds for local inference.
Release contents: - mahwizzzz/lafzyn: full weights - mahwizzzz/lafzyn-gguf: quantized builds
In 2017, Google released the Transformer architecture. While it was clear the model was promising, absolutely no one (including its authors) anticipated the pervasive global revolution it would create!
The authors actually viewed the Transformer as just a stepping stone for a much more ambitious project: The MultiModel.
Their ultimate goal was to build a single deep learning architecture capable of jointly learning massive, diverse tasks across entirely different domains (in 2017). A One Model To Learn Them All.
In fact, the MultiModel paper was published in the exact same month as Attention Is All You Need!
But history had other plans. The building block eclipsed the grand design!
So, have you heard about the MultiModel before? π
PiD β Pixel Diffusion Decoder Image Edit Upscale and Image Generation Upscale, an all-in-one demo, is now live on Spaces! Great improvements in realism-based image generation and editing are powered by FLUX.2-Klein, while image generation is paired with Z-Image, and upscaling is enabled by default!
I've made 8 Spaces in the Qwen-Image-Edit series, and out of them, 5 Spaces reached βSpace of the Weekβ! A few Spaces are still topping the list even after many months.
Cumulatively, the series has crossed 8.2 million+ ZeroGPU runs and nearly 4 million visitors overall.
Turns out : if we predict π earth we can save a lot of time looking for interesting things and less time looking at things that we expect to see.
Sentinel-2 imagery π°οΈbasically takes a long time to download towards earth. so our "near real time" systems are quite far from that in practical terms.
meanwhile , if we "predict" what we will see , based on what we do see , we can send down much less data in a timely way , and prioritize π‘earth-bound response .
I'm talking about illegal fishing , logging , mining or building in nature reserves , the more of that we predict early the more we're able to stop it on time.
Multimodal-Edge Demo, a node-based inference canvas demo, is now live on Spaces. It features node-based Transformers for fast inference across 10+ edge-device multimodal models on the Hub, all within a single space. The series includes models from Qwen3.5, Qwen3-VL, Gemma 4, and the LFM 2.5 VL model series, with support for reasoning and grounding tasks.
since everyone liked my previous announcement post ( https://huggingface.co/posts/Tonic/338509028435394 ) so much , i'm back with more high quality proceedural datasets in the Geospacial domain for SFT training !
Now, a collection of various compression schemes for Qwen3.6 and the abliterated version 1 of dense models is available on the Hub. Check it out via the links below. π
HY-World-2.0 β A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds is now available on Spaces, and it works both as native Gradio components and in Gradio server mode.
A new comparator on Spaces showcases Standard FLUX.2 Decoder vs. FLUX.2 Small Decoder. The Small Decoder is ~1.4Γ faster, uses ~1.4Γ less VRAM, and maintains near-identical image quality. It has ~28M parameters with narrower channels [96, 192, 384, 384] vs. [128, 256, 512, 512], and the demo supports sequence generation by running both decoders simultaneously and comparing the results side by side.
Now, a collection of various compression schemes for Gemma 4 and the abliterated version 1 of dense models is available on the Hub. Check it out via the links below. π
Now the demo for image detection based on SAM3 and Gemma-4 (*Filter) is available on Spaces, using full-fledged Transformers inference with multimodal reasoning for processed images. It also supports video segmentation (mask), video segmentation (annotation), and image click segmentation.
