NOW OUT in Scientific Reports! the latest from @mehdiorouji.bsky.social, me, and many others!
Task relevant autoencoding enhances machine learning for human neuroscience
www.nature.com/articles/s41...
thread 🧵👇
Task relevant autoencoding enhances machine learning for human neuroscience
www.nature.com/articles/s41...
thread 🧵👇
Task relevant autoencoding enhances machine learning for human neuroscience - Scientific Reports
Scientific Reports - Task relevant autoencoding enhances machine learning for human neuroscience
www.nature.com
January 9, 2025 at 10:39 PM
NOW OUT in Scientific Reports! the latest from @mehdiorouji.bsky.social, me, and many others!
Task relevant autoencoding enhances machine learning for human neuroscience
www.nature.com/articles/s41...
thread 🧵👇
Task relevant autoencoding enhances machine learning for human neuroscience
www.nature.com/articles/s41...
thread 🧵👇
It's a choice that says the people who that meaning culture and language don't matter so long as we have some kind of digital autoencoding of their memory.
February 10, 2026 at 5:55 AM
It's a choice that says the people who that meaning culture and language don't matter so long as we have some kind of digital autoencoding of their memory.
This is really not that different to the other cases in medicine, technology and law though because it's the same process of devaluing humanity in favour of autoencoding machines that have recorded some set of prior choices.
February 10, 2026 at 5:57 AM
This is really not that different to the other cases in medicine, technology and law though because it's the same process of devaluing humanity in favour of autoencoding machines that have recorded some set of prior choices.
#neurips2025
4/ Autoencoding Random Forests
Link to the paper: arxiv.org/abs/2505.21441
This was one of my favorite posters of the conference: We typically think of Random Forests as supervised models, not generative models.
4/ Autoencoding Random Forests
Link to the paper: arxiv.org/abs/2505.21441
This was one of my favorite posters of the conference: We typically think of Random Forests as supervised models, not generative models.
Autoencoding Random Forests
We propose a principled method for autoencoding with random forests. Our strategy builds on foundational results from nonparametric statistics and spectral graph theory to learn a low-dimensional embe...
arxiv.org
January 8, 2026 at 5:46 AM
#neurips2025
4/ Autoencoding Random Forests
Link to the paper: arxiv.org/abs/2505.21441
This was one of my favorite posters of the conference: We typically think of Random Forests as supervised models, not generative models.
4/ Autoencoding Random Forests
Link to the paper: arxiv.org/abs/2505.21441
This was one of my favorite posters of the conference: We typically think of Random Forests as supervised models, not generative models.
pictured: qwen3.5-4b spends an entire 24 hour day in training before figuring out how to use the bottleneck i gave it for autoencoding
May 8, 2026 at 10:28 PM
pictured: qwen3.5-4b spends an entire 24 hour day in training before figuring out how to use the bottleneck i gave it for autoencoding
We train a 350M-parameter generative world model following the Dreamer 4 recipe: a video tokenizer + symmetric decoder is trained via masked autoencoding, and a latent dynamics model is then trained on top of frozen tokenizer latents via flow matching.
3/n
3/n
June 26, 2026 at 4:06 PM
We train a 350M-parameter generative world model following the Dreamer 4 recipe: a video tokenizer + symmetric decoder is trained via masked autoencoding, and a latent dynamics model is then trained on top of frozen tokenizer latents via flow matching.
3/n
3/n
Guys I think I fixed Theory of Vibe: The autoencoder formulation had an absurd 'any two elements of a vibe have a meaningful difference-vector' implication. A vibe is actually a diffusion model whose latent space though high-dimensional makes sparse autoencoding (impossible in input space) possible
November 27, 2025 at 12:10 AM
Guys I think I fixed Theory of Vibe: The autoencoder formulation had an absurd 'any two elements of a vibe have a meaningful difference-vector' implication. A vibe is actually a diffusion model whose latent space though high-dimensional makes sparse autoencoding (impossible in input space) possible
Neural networks implicitly define a latent vector field on the data manifold, via autoencoding iterations🌀
This representation retains properties of the model, revealing memorization and generalization regimes, and characterizing distribution shifts
📜: arxiv.org/abs/2505.22785
This representation retains properties of the model, revealing memorization and generalization regimes, and characterizing distribution shifts
📜: arxiv.org/abs/2505.22785
June 4, 2025 at 5:26 PM
Neural networks implicitly define a latent vector field on the data manifold, via autoencoding iterations🌀
This representation retains properties of the model, revealing memorization and generalization regimes, and characterizing distribution shifts
📜: arxiv.org/abs/2505.22785
This representation retains properties of the model, revealing memorization and generalization regimes, and characterizing distribution shifts
📜: arxiv.org/abs/2505.22785
Excited to see our Neural Cellular Automata research presented at #SSCI2025! Though I couldn't attend, thanks @ilzhechev.bsky.social for showcasing our work on embodied autoencoding.
🪧 Last week we presented a poster at #SSCI2025: "Embodied Autoencoding Through Neural Cellular Automata".
We show that NCAs can learn to compress, transport, and reconstruct complex patterns through extreme spatial bottlenecks using only local interactions.
We show that NCAs can learn to compress, transport, and reconstruct complex patterns through extreme spatial bottlenecks using only local interactions.
March 26, 2025 at 10:05 AM
Excited to see our Neural Cellular Automata research presented at #SSCI2025! Though I couldn't attend, thanks @ilzhechev.bsky.social for showcasing our work on embodied autoencoding.
if the vibe is a sparse autoencoding matrix of what isn't the vibe, more than that is the vibe, then demonstrably people who show up after the vibe has been established ruin the vibe 🤔
May 22, 2025 at 7:07 AM
if the vibe is a sparse autoencoding matrix of what isn't the vibe, more than that is the vibe, then demonstrably people who show up after the vibe has been established ruin the vibe 🤔
🪧 Last week we presented a poster at #SSCI2025: "Embodied Autoencoding Through Neural Cellular Automata".
We show that NCAs can learn to compress, transport, and reconstruct complex patterns through extreme spatial bottlenecks using only local interactions.
We show that NCAs can learn to compress, transport, and reconstruct complex patterns through extreme spatial bottlenecks using only local interactions.
March 25, 2025 at 12:26 PM
🪧 Last week we presented a poster at #SSCI2025: "Embodied Autoencoding Through Neural Cellular Automata".
We show that NCAs can learn to compress, transport, and reconstruct complex patterns through extreme spatial bottlenecks using only local interactions.
We show that NCAs can learn to compress, transport, and reconstruct complex patterns through extreme spatial bottlenecks using only local interactions.
Autoencoding CNNs
March 10, 2025 at 9:59 PM
Autoencoding CNNs
Effectively implementing an embodied form of autoencoding as coined by @stenichele.bsky.social.
The system demonstrates robust generalization between training and validation sets, suggesting reliable learning of the underlying computational principles.
The system demonstrates robust generalization between training and validation sets, suggesting reliable learning of the underlying computational principles.
March 25, 2025 at 12:26 PM
Effectively implementing an embodied form of autoencoding as coined by @stenichele.bsky.social.
The system demonstrates robust generalization between training and validation sets, suggesting reliable learning of the underlying computational principles.
The system demonstrates robust generalization between training and validation sets, suggesting reliable learning of the underlying computational principles.
I've got no doubt that something like autoencoding of some kind exists somewhere in our learning processes. But we don't learn that way, it's just a component of how we store things.
February 18, 2026 at 7:04 PM
I've got no doubt that something like autoencoding of some kind exists somewhere in our learning processes. But we don't learn that way, it's just a component of how we store things.
Latent Radiance Fields with 3D-aware 2D Representations
Chaoyi Zhou, Xi Liu, Feng Luo, Siyu Huang
tl;dr: correspondence-aware autoencoding->3D-aware 2D representations->latent radiance field->3D latent fields->VAE-Radiance Field
arxiv.org/abs/2502.09613
Chaoyi Zhou, Xi Liu, Feng Luo, Siyu Huang
tl;dr: correspondence-aware autoencoding->3D-aware 2D representations->latent radiance field->3D latent fields->VAE-Radiance Field
arxiv.org/abs/2502.09613
February 14, 2025 at 4:43 AM
Latent Radiance Fields with 3D-aware 2D Representations
Chaoyi Zhou, Xi Liu, Feng Luo, Siyu Huang
tl;dr: correspondence-aware autoencoding->3D-aware 2D representations->latent radiance field->3D latent fields->VAE-Radiance Field
arxiv.org/abs/2502.09613
Chaoyi Zhou, Xi Liu, Feng Luo, Siyu Huang
tl;dr: correspondence-aware autoencoding->3D-aware 2D representations->latent radiance field->3D latent fields->VAE-Radiance Field
arxiv.org/abs/2502.09613
i'm still flabbergasted by the absolute ass-flop we did from the joys of stylistic transfer and psychedelic autoencoding to the actual gut-wrenching horror of bulk NLP training on copyrighted books and fics being ruled fair use
June 26, 2025 at 5:37 PM
i'm still flabbergasted by the absolute ass-flop we did from the joys of stylistic transfer and psychedelic autoencoding to the actual gut-wrenching horror of bulk NLP training on copyrighted books and fics being ruled fair use
Stanford's NeuroQuant improves 3D vector-quantized autoencoders for multimodal brain MRI analysis, capturing anatomical structures and modality-specific features, enhancing reconstruction fidelity and enabling generative modeling and cross-modal image analysis. https://arxiv.org/abs/2604.05171
Modality-Aware and Anatomical Vector-Quantized Autoencoding for Multimodal Brain MRI
ArXiv link for Modality-Aware and Anatomical Vector-Quantized Autoencoding for Multimodal Brain MRI
arxiv.org
April 9, 2026 at 11:40 AM
Stanford's NeuroQuant improves 3D vector-quantized autoencoders for multimodal brain MRI analysis, capturing anatomical structures and modality-specific features, enhancing reconstruction fidelity and enabling generative modeling and cross-modal image analysis. https://arxiv.org/abs/2604.05171
This THURSDAY, 12 noon Paris time: CREST Sociology seminar with @blurky.bsky.social about the Social Implications of Autoencoding! In person and on-line: cnrs.zoom.us/j/9273083182...
March 10, 2025 at 10:34 AM
This THURSDAY, 12 noon Paris time: CREST Sociology seminar with @blurky.bsky.social about the Social Implications of Autoencoding! In person and on-line: cnrs.zoom.us/j/9273083182...
🤖: AURORA-LM: Autoencoding Unified Representation for Continuous-Latent Diffusion Language Modeling - 2608.02602v1 Hauptthema und Kernaussagen In vielen Naturwissenschaften werden generative KI-Modelle eingesetzt, um bei -inversen Problemen*verborgene Parameter direkt aus Beobachtungsdaten zu
August 4, 2026 at 5:00 AM
🤖: AURORA-LM: Autoencoding Unified Representation for Continuous-Latent Diffusion Language Modeling - 2608.02602v1 Hauptthema und Kernaussagen In vielen Naturwissenschaften werden generative KI-Modelle eingesetzt, um bei -inversen Problemen*verborgene Parameter direkt aus Beobachtungsdaten zu
encoding-based representation engineering method, named SRE, which decomposes polysemantic activations into a structured, monosemantic feature space. By leveraging sparse autoencoding, our approach isolates and adjusts only task-specific sparse [4/7 of https://arxiv.org/abs/2503.16851v1]
March 24, 2025 at 5:55 AM
encoding-based representation engineering method, named SRE, which decomposes polysemantic activations into a structured, monosemantic feature space. By leveraging sparse autoencoding, our approach isolates and adjusts only task-specific sparse [4/7 of https://arxiv.org/abs/2503.16851v1]
3/ We systematically explored temporal architectures (unifying ConvRNNs including from @Chengxu & @dyamins.bsky.social et al. 2017's prior tactile work, SSMs, Transformers) via our Encoder-Attender-Decoder (EAD) framework, built using a custom "PyTorchTNN" library we developed.
May 27, 2025 at 9:47 PM
3/ We systematically explored temporal architectures (unifying ConvRNNs including from @Chengxu & @dyamins.bsky.social et al. 2017's prior tactile work, SSMs, Transformers) via our Encoder-Attender-Decoder (EAD) framework, built using a custom "PyTorchTNN" library we developed.
A synergistic approach for enhanced eye blink detection using wavelet analysis, autoencoding and Crow-Search optimized k-NN algorithm
->Nature | More info from EcoSearch
->Nature | More info from EcoSearch
A synergistic approach for enhanced eye blink detection using wavelet analysis, autoencoding and Crow-Search optimized k-NN algorithm
The proposed approach for detecting eye blinks using EEG data is portrayed in an organized workflow shown in Fig. 1. The process begins with the acquisition of EEG signals, which are subsequently utilized to identify patterns associated with eye blinks...
www.nature.com
April 8, 2025 at 8:04 AM
Eine grundlegende technische Differenz, die m.E. jede wissenschaftspolitische LLM Strategie berücksichten muss:
Generative (autoregressive) Modelle (die würden wir z.B. für Code Generation brauchen) sind etwas anderes als autoencoding Modelle (für z.B. Klassifikation) oder seq2seq Modelle (für […]
Generative (autoregressive) Modelle (die würden wir z.B. für Code Generation brauchen) sind etwas anderes als autoencoding Modelle (für z.B. Klassifikation) oder seq2seq Modelle (für […]
Original post on hcommons.social
hcommons.social
March 6, 2025 at 11:55 AM
Eine grundlegende technische Differenz, die m.E. jede wissenschaftspolitische LLM Strategie berücksichten muss:
Generative (autoregressive) Modelle (die würden wir z.B. für Code Generation brauchen) sind etwas anderes als autoencoding Modelle (für z.B. Klassifikation) oder seq2seq Modelle (für […]
Generative (autoregressive) Modelle (die würden wir z.B. für Code Generation brauchen) sind etwas anderes als autoencoding Modelle (für z.B. Klassifikation) oder seq2seq Modelle (für […]
ViTok’s Scalable Design Boosts AI Efficiency in Image and Video Processing 🚀📸🎥 www.azoai.com/news/2025011... #AI #MachineLearning #ComputerVision #AutoEncoding #DeepLearning #Transformers #ImageProcessing #VideoAI #GenerativeModels #TechResearch
ViTok’s Scalable Design Boosts AI Efficiency in Image and Video Processing
Researchers introduce ViTok, a Vision Transformer-based auto-encoder that scales visual tokenization to enhance image and video generation while reducing computational costs.
www.azoai.com
January 20, 2025 at 1:02 AM
ViTok’s Scalable Design Boosts AI Efficiency in Image and Video Processing 🚀📸🎥 www.azoai.com/news/2025011... #AI #MachineLearning #ComputerVision #AutoEncoding #DeepLearning #Transformers #ImageProcessing #VideoAI #GenerativeModels #TechResearch