Why Does A.I. Write Like … That?
Sam Kriss for the New York Times:
"""
According to the data, post-ChatGPT papers lean more on words like “underscore,” “highlight” and “showcase” than pre-ChatGPT papers [..] And “delve” [..] shot up by 2,700 percent.
"""
#EmDash #linguistics #overfitting #ElaraVoss #LLM #NYTimes #SamKriss
Researchers isolate memorization from reasoning in AI neural networks https://arstechni.ca/k2rK #mechanisticinterpretability #computationalneuroscience #AllenInstituteforAI #transformermodels #gradientdescent #machinelearning #AIarchitecture #AImemorization #generalization #neuralnetworks #weightmatrices #losscurvature #modelediting #AIalignment #overfitting #AIbehavior #AIresearch #copyright #AIsafety #Goodfire #Biz&IT #K-FAC #OLMo #AI
Eine große Fehleinschätzung ist, dass #KünstlicheNeuronaleNetzwerke umso besser werden, je komplexer sie sind und je größer der Datensatz ist, mit dem sie trainiert werden. Die aktuell völlig unterschätzte Problematik von #Overfitting & #Overtraining sind potentielle Treiber des nächsten KI-Winters. #justsaying
Nachdem hier in #Waldkirch schon seit Jahren die #Feigen keimen, habe ich jetzt auch die direkte Bestätigung, dass sich #Befruchter angesiedelt haben.
In dem Kontext musste ich meiner Mutter eine Belehrung über #Bias und #Overfitting geben, denn #KI-#Bilderkennung klassifiziert das als #Feigengallwespe. Meiner bescheidenen Laienmeinung nach handelt es sich allerdings nicht um so eine, da die Antennen m.E. völlig anders aussehen. Das System könnte also durch den Hintergrund fehlgeleitet worden sein.
Laut Wikipedia gibt es aber auch andere #Erzwespen-#Arten, die als #Feigenwespe leben. Ich würde vermuten, dass es sich um ein verwandtes #Insekt handelt, allerdings ist die FGW nach meinem Verständnis die einzige in Europa heimische Art. Also hat die KI vielleicht doch Recht.
#BlastophagaPsenes #Ficus #FicusCarica #Erzwespe #Insekten #Neophyt #Neophyten #Agaonidae
Is Grok 4 Just a Benchmark Winner? xAI's Flagship Model Raises Red Flags, Appears Heavily Overfitted to Score Well
e509 — Maverick and Marbles
e509 with Michael and Michael – stories and discussion all around #AI, #LLMs, #llamas, generated #Quake, #grokking, #generalization and much more.
https://media.blubrry.com/gamesatwork/op3.dev/e,pg=6e00562f-0386-5985-9c2c-26822923720d/gamesatwork.biz/wp-content/uploads/2025/04/E509.mp3Podcast: Play in new window | Download (Duration: 32:10 — 44.8MB) | Embed
Subscribe: Apple Podcasts | Spotify | Amazon Music | Android | Podcast Index | Youtube Music | RSS | More
Share this:
https://gamesatwork.biz/2025/04/14/e509-maverick-and-marbles/
When Dimensionality Hurts: The Role of #LLM Embedding Compression for Noisy Regression Tasks https://d.repec.org/n?u=RePEc:arx:papers:2502.02199&r=&r=cmp
"… suggest that the optimal dimensionality is dependent on the signal-to-noise ratio, exposing the necessity of feature compression in high noise environments. The implication of the result is that researchers should consider the #noise of a task when making decisions about the dimensionality of text.
… findings indicate that sentiment and emotion-based representations do not provide inherent advantages over learned latent features, implying that their previous success in similar tasks may be attributed to #regularisation effects rather than intrinsic informativeness."
#ML #autoencoders #Overfitting
'On the Impact of Hard Adversarial Instances on Overfitting in Adversarial Training', by Chen Liu, Zhichao Huang, Mathieu Salzmann, Tong Zhang, Sabine Süsstrunk.
http://jmlr.org/papers/v25/22-0950.html
#adversarial #overfitting #robustness
"Three married couples. Aren't we just boring now?" joked Mia.
"You're frowning," Ari told Jin.
"Mia's #overfitting. She's too used to seeing threats to relax to realise Tom's mum is fond of her."
"Of course she is, Tom loves Mia. How could his mum not love her too?" #vss365
1/
Recent commentary [1]:
escalating concern over the use of the more powerful #chatbots when they are used to go beyond the #knowledge of the human expert who uses them, rather than for simply pre-processing in a controlled way within the domain of human-expert knowledge.
1. What is often called "hallucination/confabulation” (i.e. severe #extrapolation #uncertainty and #overfitting by the chatbot model) is apparently becoming increasingly realistic with a declining human ability to detect it
I’m listening to #MITtechReview [PODCAST] Large language models can do jaw-dropping things. But nobody knows exactly why. (7 Aug 2024, 26 min)
https://pca.st/episode/527cdfed-17ed-4e59-8530-f9830eea78a7
#edtechSR #MediaLit #AI #lanhuage #learning #DeepLearning #magic #alchemy #OverFitting
AI image created by Wes Fryer with Ideogram:
https://ideogram.ai/g/FItzSD2BQJiJtpb7Ph0Ybw/3
🔴 Eigensolutions: composability as the antidote to overfit
by Lea Verou
@LeaVerou @leaverou
#Product #DesignThinking #CreatorTools #ProductManagement #Eigensolutions #Overfitting
'Benign Overfitting of Constant-Stepsize SGD for Linear Regression', by Difan Zou, Jingfeng Wu, Vladimir Braverman, Quanquan Gu, Sham M. Kakade.
http://jmlr.org/papers/v24/21-1297.html
#overfitting #overparameterized #sgd
🤖💡 Ever struggled with overfitting in machine learning? It can lead to poor performance and inaccurate predictions. Learn more here 👉 https://ak-codes.com/overfitting/ #MachineLearning #Overfitting
One question for the #MachineLearning people: what approach do you use to determine if a decision trees or a random forest approach should work better? Do you simply try both approaches and use whatever seems to work better?
According to what I read, decision trees are more prone to overfitting, while random forest is a more complex approach. Which means little to me 😅
Early Stopping for Deep Image Prior
Understand basic principles of underfitting and overfitting - by @dimid_ml
https://towardsdatascience.com/overfitting-and-underfitting-principles-ea8964d9c45c
#overfitting #underfitting #DataScience #MachineLearning #AI
Logistic-Normal Likelihoods for Heteroscedastic Label Noise
Erik Englesson, Amir Mehrpanah, Hossein Azizpour
Action editor: Bo Han.
Label Noise-Robust Learning using a Confidence-Based Sieving Strategy
Learning Augmentation Distributions using Transformed Risk Minimization
Evangelos Chatzipantazis, Stefanos Pertigkiozoglou, Kostas Daniilidis, Edgar Dobriban
Action editor: Andriy Mnih.
