Why do diffusion models generalise at all? It's not obvious that they would. It turns out underfitting plays an important role, as well as the architectural inductive biases of locality and translation equivariance. What other kinds of symmetry and structure could we hardcode? 🤔
Excited to finally share this work w/ @suryaganguli.bsky.social Tl;dr: we find the first closed-form analytical theory that replicates the outputs of the very simplest diffusion models, with median pixel wise r^2 values of 90%+. arxiv.org/abs/2412.20292
January 1, 2025 at 12:26 PM
Why do diffusion models generalise at all? It's not obvious that they would. It turns out underfitting plays an important role, as well as the architectural inductive biases of locality and translation equivariance. What other kinds of symmetry and structure could we hardcode? 🤔
Researchers unveil "selective underfitting" in diffusion models, showing they excel by better approximating scores in specific regions while underfitting in others, thus enhancing creative output. This insight could reshape training strategies for generative models. https://arxiv.org/abs/2510.01378
Selective Underfitting in Diffusion Models
ArXiv link for Selective Underfitting in Diffusion Models
arxiv.org
October 4, 2025 at 3:41 AM
Researchers unveil "selective underfitting" in diffusion models, showing they excel by better approximating scores in specific regions while underfitting in others, thus enhancing creative output. This insight could reshape training strategies for generative models. https://arxiv.org/abs/2510.01378
underfitting / overfitting as the two poles of our intellectual life
September 5, 2026 at 7:18 PM
underfitting / overfitting as the two poles of our intellectual life
your model has high bias and low variance. clearly underfitting :(.
February 24, 2025 at 10:37 AM
your model has high bias and low variance. clearly underfitting :(.
Evan Thomas Saitta (2026)
Are we underfitting dinosaur growth models? Accounting for intra- & inter-specific variation
Cretaceous Research 106426
doi: doi.org/10.1016/j.cr...
www.sciencedirect.com/science/arti...
Are we underfitting dinosaur growth models? Accounting for intra- & inter-specific variation
Cretaceous Research 106426
doi: doi.org/10.1016/j.cr...
www.sciencedirect.com/science/arti...
Redirecting
doi.org
May 12, 2026 at 7:35 PM
Evan Thomas Saitta (2026)
Are we underfitting dinosaur growth models? Accounting for intra- & inter-specific variation
Cretaceous Research 106426
doi: doi.org/10.1016/j.cr...
www.sciencedirect.com/science/arti...
Are we underfitting dinosaur growth models? Accounting for intra- & inter-specific variation
Cretaceous Research 106426
doi: doi.org/10.1016/j.cr...
www.sciencedirect.com/science/arti...
that one scene from whiplash except JK Simmons says "So are you overfitting or underfitting!?"
April 16, 2026 at 12:39 AM
that one scene from whiplash except JK Simmons says "So are you overfitting or underfitting!?"
I see, that the problem has something to do with what is called "underfitting" and "overfitting" in machine learning.
Some people don't learn enough to recognize fascism.
Some people learn it too detailed and don't recognize it if it does not involve uniformed Germans with hakenkreuz killing jews.
Some people don't learn enough to recognize fascism.
Some people learn it too detailed and don't recognize it if it does not involve uniformed Germans with hakenkreuz killing jews.
June 17, 2026 at 8:56 AM
I see, that the problem has something to do with what is called "underfitting" and "overfitting" in machine learning.
Some people don't learn enough to recognize fascism.
Some people learn it too detailed and don't recognize it if it does not involve uniformed Germans with hakenkreuz killing jews.
Some people don't learn enough to recognize fascism.
Some people learn it too detailed and don't recognize it if it does not involve uniformed Germans with hakenkreuz killing jews.
Yhtenä ongelmana lienee koneoppimisesta tutut "underfitting" ja "overfitting".
Underfitting: Ei tunnisteta fasismia, kun ole ajateltu tarpeeksi sitä.
Overfitting: Ei tunnisteta fasismia, koska se ei ole *täsmälleen* samassa muodossa kuin 1930-40-luvuilla.
en.wikipedia.org/wiki/Overfit...
Underfitting: Ei tunnisteta fasismia, kun ole ajateltu tarpeeksi sitä.
Overfitting: Ei tunnisteta fasismia, koska se ei ole *täsmälleen* samassa muodossa kuin 1930-40-luvuilla.
en.wikipedia.org/wiki/Overfit...
Overfitting - Wikipedia
en.wikipedia.org
July 31, 2026 at 6:39 AM
Yhtenä ongelmana lienee koneoppimisesta tutut "underfitting" ja "overfitting".
Underfitting: Ei tunnisteta fasismia, kun ole ajateltu tarpeeksi sitä.
Overfitting: Ei tunnisteta fasismia, koska se ei ole *täsmälleen* samassa muodossa kuin 1930-40-luvuilla.
en.wikipedia.org/wiki/Overfit...
Underfitting: Ei tunnisteta fasismia, kun ole ajateltu tarpeeksi sitä.
Overfitting: Ei tunnisteta fasismia, koska se ei ole *täsmälleen* samassa muodossa kuin 1930-40-luvuilla.
en.wikipedia.org/wiki/Overfit...
"how... to... fix... underfitting... with... 18... observations..."
June 6, 2026 at 8:47 PM
"how... to... fix... underfitting... with... 18... observations..."
⚠️ Friday AI Fact: Underfitting! 🔍📉
Underfitting occurs when a machine learning model is too simple to learn the underlying patterns in the data, leading to poor performance on both training and new data.
Stay tuned for more AI facts every Friday!
#ELOQUENCEAI #Underfitting #MachineLearning
Underfitting occurs when a machine learning model is too simple to learn the underlying patterns in the data, leading to poor performance on both training and new data.
Stay tuned for more AI facts every Friday!
#ELOQUENCEAI #Underfitting #MachineLearning
June 27, 2025 at 10:45 AM
⚠️ Friday AI Fact: Underfitting! 🔍📉
Underfitting occurs when a machine learning model is too simple to learn the underlying patterns in the data, leading to poor performance on both training and new data.
Stay tuned for more AI facts every Friday!
#ELOQUENCEAI #Underfitting #MachineLearning
Underfitting occurs when a machine learning model is too simple to learn the underlying patterns in the data, leading to poor performance on both training and new data.
Stay tuned for more AI facts every Friday!
#ELOQUENCEAI #Underfitting #MachineLearning
Finding an interesting theorem as an interplay between avoiding vacuously true statements and ???¹ Example of Abduction
¹: Something about the interplay between overfitting and underfitting?
¹: Something about the interplay between overfitting and underfitting?
March 13, 2026 at 11:56 PM
Finding an interesting theorem as an interplay between avoiding vacuously true statements and ???¹ Example of Abduction
¹: Something about the interplay between overfitting and underfitting?
¹: Something about the interplay between overfitting and underfitting?
By introducing more and more complexity, and by requiring more and more set parameters, any such model will eventually collapse into total gibberish or direct reconstruction. 'Underfitting' and 'overfitting' aren't simply bugs but inherent vices
March 20, 2025 at 6:05 PM
By introducing more and more complexity, and by requiring more and more set parameters, any such model will eventually collapse into total gibberish or direct reconstruction. 'Underfitting' and 'overfitting' aren't simply bugs but inherent vices
It's not the code - it's the refining of the model. The classic overfitting/underfitting problems persist as well as black box model analysis. It's often not profitable to create a one-task ML model that does one thing really well.
September 28, 2026 at 7:05 PM
It's not the code - it's the refining of the model. The classic overfitting/underfitting problems persist as well as black box model analysis. It's often not profitable to create a one-task ML model that does one thing really well.
[..]creating a model that performs well on unseen data - generalizing beyond the training set - is one of the central challenges in ML. This challenge often revolves around two key issues: overfitting and underfitting[..]
#machine #learning #model #ai
www.ml-nn.eu/a1/52.html
#machine #learning #model #ai
www.ml-nn.eu/a1/52.html
Overfitting and Underfitting in Machine Learning
Machine Learning & Neural Networks Blog
www.ml-nn.eu
July 2, 2026 at 10:41 AM
[..]creating a model that performs well on unseen data - generalizing beyond the training set - is one of the central challenges in ML. This challenge often revolves around two key issues: overfitting and underfitting[..]
#machine #learning #model #ai
www.ml-nn.eu/a1/52.html
#machine #learning #model #ai
www.ml-nn.eu/a1/52.html
Step back and ask yourself the questions out loud about what you're seeing and how it either does or does not make sense instead of underfitting people into weird arbitrary binaries.
Again.
Again.
July 7, 2026 at 1:00 PM
Step back and ask yourself the questions out loud about what you're seeing and how it either does or does not make sense instead of underfitting people into weird arbitrary binaries.
Again.
Again.
Building accurate and reliable machine learning models is not without its challenges. One critical aspect that demands careful consideration is striking the delicate balance between model complexity and generalization. #ML #Overfitting #Underfitting #MachineLearning
Understanding Overfitting and Underfitting in Machine Learning
Machine learning has become an integral part of our lives, powering applications and technologies that range from personalized…
medium.com
March 16, 2024 at 3:23 AM
Building accurate and reliable machine learning models is not without its challenges. One critical aspect that demands careful consideration is striking the delicate balance between model complexity and generalization. #ML #Overfitting #Underfitting #MachineLearning
Day317 youtu.be/o3DztvnfAJg ai overfitting #100daysofcode
Underfitting & Overfitting - Explained
YouTube video by NStatum
youtu.be
May 2, 2026 at 4:25 PM
Day317 youtu.be/o3DztvnfAJg ai overfitting #100daysofcode
It is super cool to hear many many views and compare them. It doesn't become "argument to moderation" (false middle) but instead prevents underfitting and means I get a view with real depth.
August 20, 2025 at 4:52 PM
It is super cool to hear many many views and compare them. It doesn't become "argument to moderation" (false middle) but instead prevents underfitting and means I get a view with real depth.
Overfitting vs. Underfitting: Making Sense of the Bias-Variance Trade-Off
The best models live in the sweet spot: generalizing well, learning enough, but not too much
Telegram AI Digest
#ai #overfitting #underfitting
The best models live in the sweet spot: generalizing well, learning enough, but not too much
Telegram AI Digest
#ai #overfitting #underfitting
Overfitting vs. Underfitting: Making Sense of the Bias-Variance Trade-Off
The best models live in the sweet spot: generalizing well, learning enough, but not too much
towardsdatascience.com
November 23, 2025 at 9:08 AM
Overfitting vs. Underfitting: Making Sense of the Bias-Variance Trade-Off
The best models live in the sweet spot: generalizing well, learning enough, but not too much
Telegram AI Digest
#ai #overfitting #underfitting
The best models live in the sweet spot: generalizing well, learning enough, but not too much
Telegram AI Digest
#ai #overfitting #underfitting
One of my students who is studying LLM performance for his masters is convinced that the entire way these products work is based on overfitting, and that the reason this isn’t widely’ recognised is because the models are so colossal in size that it’s harder to see directly than in traditional ML.
June 15, 2025 at 1:11 AM
One of my students who is studying LLM performance for his masters is convinced that the entire way these products work is based on overfitting, and that the reason this isn’t widely’ recognised is because the models are so colossal in size that it’s harder to see directly than in traditional ML.
⚖️ El Equilibrio del Machine Learning
Sesgo alto = underfitting. Varianza alta = overfitting. El punto óptimo combina bajo sesgo y baja varianza. Y entre dos modelos iguales, el más simple gana.
#MachineLearning #SesgoVarianza #Overfitting #Underfitting #NavajaDeOckham #CienciaDeDatos
Sesgo alto = underfitting. Varianza alta = overfitting. El punto óptimo combina bajo sesgo y baja varianza. Y entre dos modelos iguales, el más simple gana.
#MachineLearning #SesgoVarianza #Overfitting #Underfitting #NavajaDeOckham #CienciaDeDatos
September 22, 2026 at 8:05 PM
⚖️ El Equilibrio del Machine Learning
Sesgo alto = underfitting. Varianza alta = overfitting. El punto óptimo combina bajo sesgo y baja varianza. Y entre dos modelos iguales, el más simple gana.
#MachineLearning #SesgoVarianza #Overfitting #Underfitting #NavajaDeOckham #CienciaDeDatos
Sesgo alto = underfitting. Varianza alta = overfitting. El punto óptimo combina bajo sesgo y baja varianza. Y entre dos modelos iguales, el más simple gana.
#MachineLearning #SesgoVarianza #Overfitting #Underfitting #NavajaDeOckham #CienciaDeDatos
The “bias-variance tradeoff” helps us understand how our model is working: If a model has high bias, it does poorly on both training data and test data, and if a model has high variance, it does very well on training data but poorly on test data. #ML #AI #AI
Bias-Variance Tradeoff, Explained: A Visual Guide with Code Examples for Beginners
How underfitting and overfitting fight over your models
towardsdatascience.com
December 10, 2024 at 2:41 AM
* weird brain disease. adhd hunters and schizo shamans. comorbidity ball? boat dream. acute overfitting and chronic underfitting. psychopath army climber guy story
* precarity and power law. comparative advantage, serfs/wages, gig/workshop. swing, how to own a person, kodak vs apple
* precarity and power law. comparative advantage, serfs/wages, gig/workshop. swing, how to own a person, kodak vs apple
May 7, 2023 at 8:34 AM
* weird brain disease. adhd hunters and schizo shamans. comorbidity ball? boat dream. acute overfitting and chronic underfitting. psychopath army climber guy story
* precarity and power law. comparative advantage, serfs/wages, gig/workshop. swing, how to own a person, kodak vs apple
* precarity and power law. comparative advantage, serfs/wages, gig/workshop. swing, how to own a person, kodak vs apple
Don’t fear underfitting early.
Overfitting is the real monster.
Dropout = silver bullet.
#100DaysOfCode #AI
#MachineLearning
Overfitting is the real monster.
Dropout = silver bullet.
#100DaysOfCode #AI
#MachineLearning
August 2, 2025 at 5:13 PM
Don’t fear underfitting early.
Overfitting is the real monster.
Dropout = silver bullet.
#100DaysOfCode #AI
#MachineLearning
Overfitting is the real monster.
Dropout = silver bullet.
#100DaysOfCode #AI
#MachineLearning