Convolutional Differentiable Logic Gate Networks @FHKPetersen
December 13, 2024 at 8:24 PM
Convolutional Differentiable Logic Gate Networks @FHKPetersen
New video: DLSS 4's transformer model shows dramatic improvements over the older convolutional neural network version - and it's best seen in ray reconstruction, as Alex reports: youtu.be/rlePeTM-tv0
January 25, 2025 at 3:46 PM
New video: DLSS 4's transformer model shows dramatic improvements over the older convolutional neural network version - and it's best seen in ray reconstruction, as Alex reports: youtu.be/rlePeTM-tv0
🚨 New preprint alert! We trained a convolutional neural network to predict an observer's task by guiding each convolutional layer with EEG information. Different time points guided different layers. #neuroskyence 1/
January 3, 2025 at 6:38 PM
🚨 New preprint alert! We trained a convolutional neural network to predict an observer's task by guiding each convolutional layer with EEG information. Different time points guided different layers. #neuroskyence 1/
it has been a trip watching convolutional methods fall from golden child status
September 23, 2026 at 4:13 PM
it has been a trip watching convolutional methods fall from golden child status
i think what's going on is something like that: the aggregate representation of some feature too large to represent well in the small convolutional blocks forms a representation in a larger convolutional block, whereupon internal covariate shift problems breaks the small convolutional blocks.
Maybe it continues to compress existing information stored in the weights to make room for new information even as its not really getting any better at producing the final result?
October 22, 2024 at 7:28 PM
i think what's going on is something like that: the aggregate representation of some feature too large to represent well in the small convolutional blocks forms a representation in a larger convolutional block, whereupon internal covariate shift problems breaks the small convolutional blocks.
Updated preprint and code from my lab. If you want to explore more realistic neural dynamics in models that can still perform visual tasks, check it out!
www.biorxiv.org/content/10.1...
www.biorxiv.org/content/10.1...
Modeling Dynamical Vision with Biologically Plausible Recurrent Convolutional Networks
Convolutional Neural Networks (CNNs) trained for image recognition have demonstrated remarkable conceptual similarities to the primate ventral visual pathway, but their standard feedforward architectu...
www.biorxiv.org
June 26, 2026 at 4:42 PM
Updated preprint and code from my lab. If you want to explore more realistic neural dynamics in models that can still perform visual tasks, check it out!
www.biorxiv.org/content/10.1...
www.biorxiv.org/content/10.1...
My #NeurIPS204 workshop keynote: "Analytic theory of creativity of convolutional diffusion models" now available: neurips.cc/virtual/2024...
It reports on work with @MasonKamb in this paper: arxiv.org/abs/2412.20292
It reports on work with @MasonKamb in this paper: arxiv.org/abs/2412.20292
NeurIPS Surya Ganguli: An analytic theory of creativity in convolutional diffusion modelsNeurIPS 2024
neurips.cc
January 24, 2025 at 3:19 PM
My #NeurIPS204 workshop keynote: "Analytic theory of creativity of convolutional diffusion models" now available: neurips.cc/virtual/2024...
It reports on work with @MasonKamb in this paper: arxiv.org/abs/2412.20292
It reports on work with @MasonKamb in this paper: arxiv.org/abs/2412.20292
This is an incredible visualisation of Convolutional neural network
Source in comments:
Source in comments:
November 27, 2024 at 3:07 PM
This is an incredible visualisation of Convolutional neural network
Source in comments:
Source in comments:
was fairly obvious that people in the long run are incapable of separating gAI from stuff like convolutional neural networks on images, which basically do incredible things
Despise more than anything that people assume all AI can have hallucinations now
This comment (and ensuing discussion) are really good in particular. I think everyone here has a point (with the third comment aligning more or less with how I see things), but the fascinating part for me is that the more visibility the general public has with AI, the worse its reputation becomes
June 27, 2025 at 7:23 PM
was fairly obvious that people in the long run are incapable of separating gAI from stuff like convolutional neural networks on images, which basically do incredible things
So, noch ein Thread zu Neural Networks bevor ich mehr zu Convolutional Neural Networks sagen kann. 🧵
January 7, 2026 at 4:57 PM
So, noch ein Thread zu Neural Networks bevor ich mehr zu Convolutional Neural Networks sagen kann. 🧵
Our new paper! "Analytic theory of creativity in convolutional diffusion models" lead expertly by @masonkamb.bsky.social
arxiv.org/abs/2412.20292
Our closed-form theory needs no training, is mechanistically interpretable & accurately predicts diffusion model outputs with high median r^2~0.9
arxiv.org/abs/2412.20292
Our closed-form theory needs no training, is mechanistically interpretable & accurately predicts diffusion model outputs with high median r^2~0.9
December 31, 2024 at 4:54 PM
Our new paper! "Analytic theory of creativity in convolutional diffusion models" lead expertly by @masonkamb.bsky.social
arxiv.org/abs/2412.20292
Our closed-form theory needs no training, is mechanistically interpretable & accurately predicts diffusion model outputs with high median r^2~0.9
arxiv.org/abs/2412.20292
Our closed-form theory needs no training, is mechanistically interpretable & accurately predicts diffusion model outputs with high median r^2~0.9
the team at rhizomatiks handled the hardware, visualization, & data collection. i handled blade tracking, which meant training multiple custom convolutional neural networks and deploying them across 24 GPUs to track the entire curve of the 1px wide weapons at 60fps in realtime
April 20, 2026 at 8:26 PM
the team at rhizomatiks handled the hardware, visualization, & data collection. i handled blade tracking, which meant training multiple custom convolutional neural networks and deploying them across 24 GPUs to track the entire curve of the 1px wide weapons at 60fps in realtime
By 2022 Google Translate had been using transformer models for several years.
By 2014 Facebook was using deep convolutional neural nets (DeepFave) for face detection.
Literally the things you think were “before LLMs” are either basically indistinguishable from LLMs or their direct predecessors.
By 2014 Facebook was using deep convolutional neural nets (DeepFave) for face detection.
Literally the things you think were “before LLMs” are either basically indistinguishable from LLMs or their direct predecessors.
Right but I had contractors using these translation apps with me in 2022 before they were made of large models. Facebook was auto-tagging people in photos a decade ago (contrary to the QT suggesting computer vision is new). Lots of “AI” things that are not based on large models are def useful.
August 2, 2026 at 4:07 PM
By 2022 Google Translate had been using transformer models for several years.
By 2014 Facebook was using deep convolutional neural nets (DeepFave) for face detection.
Literally the things you think were “before LLMs” are either basically indistinguishable from LLMs or their direct predecessors.
By 2014 Facebook was using deep convolutional neural nets (DeepFave) for face detection.
Literally the things you think were “before LLMs” are either basically indistinguishable from LLMs or their direct predecessors.
owo what's this *punctures your convolutional code*
November 30, 2025 at 4:29 AM
owo what's this *punctures your convolutional code*
damn convolutional neural network image denoising is so cool
March 20, 2024 at 6:28 PM
damn convolutional neural network image denoising is so cool
August 4, 2025 at 12:33 PM
yes, it's such a huge departure from the immediately preceding themes like "the convolutional layers are exactly how the visual cortex is structured" or "we built the recurrence mechanism like a tape-marking machine, so it's Turing complete." nope. attentions is all you need :(
September 23, 2026 at 4:03 PM
yes, it's such a huge departure from the immediately preceding themes like "the convolutional layers are exactly how the visual cortex is structured" or "we built the recurrence mechanism like a tape-marking machine, so it's Turing complete." nope. attentions is all you need :(
A great @quantamagazine.bsky.social article on our theory of creativity in convolutional diffusion models lead by @masonkamb.bsky.social See also our paper with new results in version 2: arxiv.org/abs/2412.20292 to be presented as an oral at @icmlconf.bsky.social #icml25
June 30, 2025 at 6:05 PM
A great @quantamagazine.bsky.social article on our theory of creativity in convolutional diffusion models lead by @masonkamb.bsky.social See also our paper with new results in version 2: arxiv.org/abs/2412.20292 to be presented as an oral at @icmlconf.bsky.social #icml25
From pixels to predictions - this is how AI sees! 👀🤖
A 𝗖𝗼𝗻𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻𝗮𝗹 𝗡𝗲𝘂𝗿𝗮𝗹 𝗡𝗲𝘁𝘄𝗼𝗿𝗸 breaks down images step by step:
🔹 Convolutional Layers → Detect edges, textures & patterns
🔹 Pooling Layers → Shrink images while keeping key info
🔹 Fully Connected Layers → Weigh features & make final decisions
A 𝗖𝗼𝗻𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻𝗮𝗹 𝗡𝗲𝘂𝗿𝗮𝗹 𝗡𝗲𝘁𝘄𝗼𝗿𝗸 breaks down images step by step:
🔹 Convolutional Layers → Detect edges, textures & patterns
🔹 Pooling Layers → Shrink images while keeping key info
🔹 Fully Connected Layers → Weigh features & make final decisions
March 4, 2025 at 1:50 PM
From pixels to predictions - this is how AI sees! 👀🤖
A 𝗖𝗼𝗻𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻𝗮𝗹 𝗡𝗲𝘂𝗿𝗮𝗹 𝗡𝗲𝘁𝘄𝗼𝗿𝗸 breaks down images step by step:
🔹 Convolutional Layers → Detect edges, textures & patterns
🔹 Pooling Layers → Shrink images while keeping key info
🔹 Fully Connected Layers → Weigh features & make final decisions
A 𝗖𝗼𝗻𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻𝗮𝗹 𝗡𝗲𝘂𝗿𝗮𝗹 𝗡𝗲𝘁𝘄𝗼𝗿𝗸 breaks down images step by step:
🔹 Convolutional Layers → Detect edges, textures & patterns
🔹 Pooling Layers → Shrink images while keeping key info
🔹 Fully Connected Layers → Weigh features & make final decisions
Let me tell you about some other random shit while we wait for the model to train: Did you know that almost all convolutional architectures have a flaw that makes them suboptimal for segmentation or other 2d tasks?
It's the padding! Let me show you how to fix it!🧵 #mlsky
It's the padding! Let me show you how to fix it!🧵 #mlsky
November 19, 2024 at 12:26 PM
Let me tell you about some other random shit while we wait for the model to train: Did you know that almost all convolutional architectures have a flaw that makes them suboptimal for segmentation or other 2d tasks?
It's the padding! Let me show you how to fix it!🧵 #mlsky
It's the padding! Let me show you how to fix it!🧵 #mlsky
The article doesn't say it, but this cool solar forecasting software uses no large language models, the thing people usually think of when they hear "AI." It uses a fairly simple convolutional neural network that's readily trained on a laptop. No data centers or LLMs involved. Research paper here:
November 9, 2025 at 12:29 PM
The article doesn't say it, but this cool solar forecasting software uses no large language models, the thing people usually think of when they hear "AI." It uses a fairly simple convolutional neural network that's readily trained on a laptop. No data centers or LLMs involved. Research paper here:
Want to check out the source for the "AlexNet" paper? Google has made the code from Krizhevsky, Sutskever and Hinton's seminal "ImageNet Classification with Deep Convolutional
Neural Networks" paper open source, in partnership with the Computer History Museum.
computerhistory.org/press-releas...
Neural Networks" paper open source, in partnership with the Computer History Museum.
computerhistory.org/press-releas...
March 20, 2025 at 9:02 PM
Want to check out the source for the "AlexNet" paper? Google has made the code from Krizhevsky, Sutskever and Hinton's seminal "ImageNet Classification with Deep Convolutional
Neural Networks" paper open source, in partnership with the Computer History Museum.
computerhistory.org/press-releas...
Neural Networks" paper open source, in partnership with the Computer History Museum.
computerhistory.org/press-releas...
it's the algorithm that's behind LLMs and modern image/video generators and ~most things that get lumped under "AI" these days, invented in 2017. arxiv.org/abs/1706.03762
Attention Is All You Need
The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and d...
arxiv.org
September 25, 2026 at 3:46 PM
it's the algorithm that's behind LLMs and modern image/video generators and ~most things that get lumped under "AI" these days, invented in 2017. arxiv.org/abs/1706.03762
yet another situation where a simple markov chain would be the best option instead of a convolutional neural net, but some people just want to make things complicated for no reason
March 17, 2025 at 11:04 PM
yet another situation where a simple markov chain would be the best option instead of a convolutional neural net, but some people just want to make things complicated for no reason