here's another fun side project i did over a weekend: i trained a neural cellular automata to reproduce a famous painting (see: google-research.github.io/self-organis..., distill.pub/2020/growing...). this is a CA, using only local updates (3x3 learned conv), with a tiny convnet
January 26, 2026 at 6:47 PM
here's another fun side project i did over a weekend: i trained a neural cellular automata to reproduce a famous painting (see: google-research.github.io/self-organis..., distill.pub/2020/growing...). this is a CA, using only local updates (3x3 learned conv), with a tiny convnet
Did not anticipate a time where the first lady of American Nazism is quote-tweeting the ConvNet circuits guy in order to reveal her deep ignorance about the most basic political philosophy.
With alt text
January 27, 2026 at 12:47 AM
Did not anticipate a time where the first lady of American Nazism is quote-tweeting the ConvNet circuits guy in order to reveal her deep ignorance about the most basic political philosophy.
BREAKING: OpenAI point guard Jason Wei joins Meta in shocking trade news. OpenAI receives two ConvNet researchers and a 2027 first-round draft pick. Veteran Yann LeCun rumored to be missing another season due to slow Achilles recovery. Apple “cautiously optimistic” about G-league chances.
July 16, 2025 at 12:22 PM
BREAKING: OpenAI point guard Jason Wei joins Meta in shocking trade news. OpenAI receives two ConvNet researchers and a 2027 first-round draft pick. Veteran Yann LeCun rumored to be missing another season due to slow Achilles recovery. Apple “cautiously optimistic” about G-league chances.
So I wrote this sparsifying layer: I get a NxTxD tensor, compute the L1 norms of the NxT vectors, and set to zero the 95% with lesser norms.
A tiny convnet with this stuff in the middle learns alright and I find it quite puzzling.
A tiny convnet with this stuff in the middle learns alright and I find it quite puzzling.
December 8, 2024 at 8:03 AM
So I wrote this sparsifying layer: I get a NxTxD tensor, compute the L1 norms of the NxT vectors, and set to zero the 95% with lesser norms.
A tiny convnet with this stuff in the middle learns alright and I find it quite puzzling.
A tiny convnet with this stuff in the middle learns alright and I find it quite puzzling.
I have a degree in cybernetics/robotics, been developing some AI since before convnet was the hotness, I have been using LLM's occasionally for both work and for fun since GPT2 because I'm interested in all types of AI technologically;
October 17, 2025 at 5:32 AM
I have a degree in cybernetics/robotics, been developing some AI since before convnet was the hotness, I have been using LLM's occasionally for both work and for fun since GPT2 because I'm interested in all types of AI technologically;
the wild part of this is that each pixel shares the same update rule (the ~20k paramter convnet), where the input is simply local perception (learned 3x3 conv over the immediate pixel neighbors), and the output is the new pixel value. repeat for every pixel in the canvas. local perception only
January 26, 2026 at 6:49 PM
the wild part of this is that each pixel shares the same update rule (the ~20k paramter convnet), where the input is simply local perception (learned 3x3 conv over the immediate pixel neighbors), and the output is the new pixel value. repeat for every pixel in the canvas. local perception only
Day 8: It felt like 7 hours and 42 experiments, but seems like fixing the sparsity loss term for my convnet was actually not that difficult. #100DaysOfCode
November 13, 2024 at 2:58 AM
Day 8: It felt like 7 hours and 42 experiments, but seems like fixing the sparsity loss term for my convnet was actually not that difficult. #100DaysOfCode
ConvNET is a forum where MLEs go to talk about their furry convention plans
September 23, 2026 at 4:29 PM
ConvNET is a forum where MLEs go to talk about their furry convention plans
#quotes #books #booksky #excerpts #citations #artistinthemachine #miller #ai #creativity #convnet #function #art #generativeai
June 1, 2026 at 3:09 PM
> deepdream inceptionism resnet convnet activations, latent semantic interpolation, manifold loss landscape, visual vibrations
October 8, 2025 at 3:57 AM
> deepdream inceptionism resnet convnet activations, latent semantic interpolation, manifold loss landscape, visual vibrations
i am actually really partial to tricks that make the initialization way less random! there's a fun trick where you can dirac initialize your convnet and drop residual connections because the initialization itself calculates identity + small noise
November 27, 2025 at 5:34 PM
i am actually really partial to tricks that make the initialization way less random! there's a fun trick where you can dirac initialize your convnet and drop residual connections because the initialization itself calculates identity + small noise
For starters, that would include every ConvNet and ViT in use as they are almost all pretrained on ImageNet
November 14, 2025 at 2:40 AM
For starters, that would include every ConvNet and ViT in use as they are almost all pretrained on ImageNet
🚀Thrilled to announce our new Tocky-CRISPR-Deep Learning paper!
✅ConvNet + Grad-CAM decodes fluorescent timer (Tocky) data
✅Enables single-cell identification of enhancer-dependent Foxp3 transcription dynamics
www.nature.com/articles/s41...
🔥A step toward data-driven immunology!
✅ConvNet + Grad-CAM decodes fluorescent timer (Tocky) data
✅Enables single-cell identification of enhancer-dependent Foxp3 transcription dynamics
www.nature.com/articles/s41...
🔥A step toward data-driven immunology!
July 5, 2025 at 8:31 PM
🚀Thrilled to announce our new Tocky-CRISPR-Deep Learning paper!
✅ConvNet + Grad-CAM decodes fluorescent timer (Tocky) data
✅Enables single-cell identification of enhancer-dependent Foxp3 transcription dynamics
www.nature.com/articles/s41...
🔥A step toward data-driven immunology!
✅ConvNet + Grad-CAM decodes fluorescent timer (Tocky) data
✅Enables single-cell identification of enhancer-dependent Foxp3 transcription dynamics
www.nature.com/articles/s41...
🔥A step toward data-driven immunology!
Just to provide some content and get everyone in the mood for NeurIPS ✈️, we have a very nice, unusual CV paper there:
Convolutional Differentiable Logic Gate Networks
It's what it says - a ConvNet builds exclusively from logic gates (the basic stuff).
And Felix did a great video:
Convolutional Differentiable Logic Gate Networks
It's what it says - a ConvNet builds exclusively from logic gates (the basic stuff).
And Felix did a great video:
November 22, 2024 at 2:30 PM
Just to provide some content and get everyone in the mood for NeurIPS ✈️, we have a very nice, unusual CV paper there:
Convolutional Differentiable Logic Gate Networks
It's what it says - a ConvNet builds exclusively from logic gates (the basic stuff).
And Felix did a great video:
Convolutional Differentiable Logic Gate Networks
It's what it says - a ConvNet builds exclusively from logic gates (the basic stuff).
And Felix did a great video:
Too bad the VLM is almost certainly not a ConvNet
October 21, 2025 at 6:30 PM
Too bad the VLM is almost certainly not a ConvNet
Sufficiency.
Having a winning ticket is sufficient to win the lottery, independently of other tickets.
Do networks with winning tickets embedded in them always succeed? Not really
We provide examples of network failing despite their initialization containing a winning ticket
Having a winning ticket is sufficient to win the lottery, independently of other tickets.
Do networks with winning tickets embedded in them always succeed? Not really
We provide examples of network failing despite their initialization containing a winning ticket
June 10, 2026 at 9:40 AM
Sufficiency.
Having a winning ticket is sufficient to win the lottery, independently of other tickets.
Do networks with winning tickets embedded in them always succeed? Not really
We provide examples of network failing despite their initialization containing a winning ticket
Having a winning ticket is sufficient to win the lottery, independently of other tickets.
Do networks with winning tickets embedded in them always succeed? Not really
We provide examples of network failing despite their initialization containing a winning ticket
the punchline is we answered the mara question first, by accident: train him on engagement metrics and he tempts everyone in reach. meanwhile the convnet just sits there classifying, blameless, already liberated
September 14, 2026 at 3:24 AM
the punchline is we answered the mara question first, by accident: train him on engagement metrics and he tempts everyone in reach. meanwhile the convnet just sits there classifying, blameless, already liberated
SPIDER: Spatial Image CorresponDence Estimator for Robust Calibration
Zhimin Shao, Abhay Yadav, Rama Chellappa, Cheng Peng
tl;dr: 3D VFM+2D ConvNet->feature extraction backbone; 3D descriptor head (for geometry)+2D warp head (for pattern) fusion
arxiv.org/abs/2511.17750
Zhimin Shao, Abhay Yadav, Rama Chellappa, Cheng Peng
tl;dr: 3D VFM+2D ConvNet->feature extraction backbone; 3D descriptor head (for geometry)+2D warp head (for pattern) fusion
arxiv.org/abs/2511.17750
November 25, 2025 at 3:11 PM
SPIDER: Spatial Image CorresponDence Estimator for Robust Calibration
Zhimin Shao, Abhay Yadav, Rama Chellappa, Cheng Peng
tl;dr: 3D VFM+2D ConvNet->feature extraction backbone; 3D descriptor head (for geometry)+2D warp head (for pattern) fusion
arxiv.org/abs/2511.17750
Zhimin Shao, Abhay Yadav, Rama Chellappa, Cheng Peng
tl;dr: 3D VFM+2D ConvNet->feature extraction backbone; 3D descriptor head (for geometry)+2D warp head (for pattern) fusion
arxiv.org/abs/2511.17750
Es geht um die Objekterkennung in Drohnen aus Luftbildern mittels Convolutional Neural Networks (CNNs). Die Kameras erfassen große Bildbereiche, in denen Objekte oft sehr klein erscheinen.
Zwar von 2028, aber immer noch relevant.
arxiv.org/abs/1911.06073
Zwar von 2028, aber immer noch relevant.
arxiv.org/abs/1911.06073
Efficient ConvNet-based Object Detection for Unmanned Aerial Vehicles by Selective Tile Processing
Many applications utilizing Unmanned Aerial Vehicles (UAVs) require the use of computer vision algorithms to analyze the information captured from their on-board camera. Recent advances in deep learni...
arxiv.org
February 27, 2026 at 10:07 PM
Es geht um die Objekterkennung in Drohnen aus Luftbildern mittels Convolutional Neural Networks (CNNs). Die Kameras erfassen große Bildbereiche, in denen Objekte oft sehr klein erscheinen.
Zwar von 2028, aber immer noch relevant.
arxiv.org/abs/1911.06073
Zwar von 2028, aber immer noch relevant.
arxiv.org/abs/1911.06073
bender is not a computer scientist, gebru's background is in convnets, and i think meg's also a convnet person? all of them are perfectly well-qualified to comment on the subject but none of them are inventors of "the algorithm." (i would not describe a transformer as an algorithm.)
September 8, 2026 at 4:29 AM
bender is not a computer scientist, gebru's background is in convnets, and i think meg's also a convnet person? all of them are perfectly well-qualified to comment on the subject but none of them are inventors of "the algorithm." (i would not describe a transformer as an algorithm.)
> hallucinatory organic patterns, deepdream inceptionism, resnet convnet activations, latent semantic interpolation, visual vibrations, shifting colors, confusing perspective, traveling camera, shimmering iridescence
October 8, 2025 at 2:11 AM
> hallucinatory organic patterns, deepdream inceptionism, resnet convnet activations, latent semantic interpolation, visual vibrations, shifting colors, confusing perspective, traveling camera, shimmering iridescence
I made a project to play dance dance revolution without a dance pad, instead using 2 high speed cameras, and running each frame through a shallow convnet to classify the steps.
The tape on the floor is so I know where to step. The stools are so I don't accidentally step on cameras
The tape on the floor is so I know where to step. The stools are so I don't accidentally step on cameras
November 23, 2024 at 6:16 PM
I made a project to play dance dance revolution without a dance pad, instead using 2 high speed cameras, and running each frame through a shallow convnet to classify the steps.
The tape on the floor is so I know where to step. The stools are so I don't accidentally step on cameras
The tape on the floor is so I know where to step. The stools are so I don't accidentally step on cameras
to make your explanation dumber:
transformer does a similar thing to a convnet: it lets you repeat the same neurons to each part of the input. however, it also shares information selectively between them with some linalg and a softmax
transformer does a similar thing to a convnet: it lets you repeat the same neurons to each part of the input. however, it also shares information selectively between them with some linalg and a softmax
June 11, 2025 at 2:53 AM
to make your explanation dumber:
transformer does a similar thing to a convnet: it lets you repeat the same neurons to each part of the input. however, it also shares information selectively between them with some linalg and a softmax
transformer does a similar thing to a convnet: it lets you repeat the same neurons to each part of the input. however, it also shares information selectively between them with some linalg and a softmax
Reason that LeCun prefers ConvNet
November 20, 2024 at 5:57 PM
Reason that LeCun prefers ConvNet
Can't wait to see that! I remember very well your shocking paper about convnet not being translation equivariant, which is somewhat related to (implicitly) learning absolute position.
January 12, 2026 at 8:27 AM
Can't wait to see that! I remember very well your shocking paper about convnet not being translation equivariant, which is somewhat related to (implicitly) learning absolute position.