New AI Technique Removes Hidden Bias Without Needing to Know What It Is

[konerto]

Artificial intelligence models often fall into the trap of relying on “spurious correlations”—patterns in training data that are irrelevant or misleading, but which the AI uses to make decisions. Now, researchers from North Carolina State University have introduced a new method that can eliminate these biases, even when the exact source of the problem is unknown. Their findings are available on the arXiv preprint server and will be presented at the upcoming International Conference on Learning Representations (ICLR) in Singapore.

“What's novel about our technique is that it works even when you don’t know what spurious features the AI is depending on,” said Jung-Eun Kim, corresponding author of the study and assistant professor of computer science at NC State.

Spurious correlations often arise due to what’s known as simplicity bias in AI training. For instance, if an AI model is trained to identify dogs from photos, and many dogs in the training set are wearing collars, the model might begin to associate collars—not actual dog features like ears or fur—with dogs. As a result, it might incorrectly identify cats wearing collars as dogs.

Traditional methods for fixing this issue typically require that practitioners know which features are spurious and then modify the dataset accordingly—such as by including more images of dogs without collars. But in many cases, identifying the misleading feature isn’t possible, making those methods ineffective.

To address this, Kim and her team developed a data-pruning approach that doesn’t require prior knowledge of spurious features. The technique works by removing a small subset of the most "difficult" samples in the training dataset—those that cause the model the most uncertainty or confusion during learning.

https://www.ediogames.com/fr/forum/topic/3290/1
https://www.ediogames.com/fr/forum/topic/3291/1
https://www.ediogames.com/fr/forum/topic/3292/1
https://www.ediogames.com/fr/forum/topic/3293/1
https://www.ediogames.com/fr/forum/topic/3294/1
https://www.ediogames.com/fr/forum/topic/3295/1
https://www.ediogames.com/fr/forum/topic/3296/1
https://www.ediogames.com/fr/forum/topic/3297/1
https://www.ediogames.com/fr/forum/topic/3307/1
https://www.ediogames.com/fr/forum/topic/3306/1
https://www.ediogames.com/fr/forum/topic/3305/1
https://www.ediogames.com/fr/forum/topic/3304/1
https://www.ediogames.com/fr/forum/topic/3303/1
https://www.ediogames.com/fr/forum/topic/3302/1
https://www.ediogames.com/fr/forum/topic/3301/1
https://www.ediogames.com/fr/forum/topic/3300/1
https://www.ediogames.com/fr/forum/topic/3299/1
https://www.ediogames.com/fr/forum/topic/3298/1
https://www.ediogames.com/fr/forum/topic/3317/1
https://www.ediogames.com/fr/forum/topic/3316/1
https://www.ediogames.com/fr/forum/topic/3315/1
https://www.ediogames.com/fr/forum/topic/3314/1
https://www.ediogames.com/fr/forum/topic/3313/1
https://www.ediogames.com/fr/forum/topic/3312/1
https://www.ediogames.com/fr/forum/topic/3311/1
https://www.ediogames.com/fr/forum/topic/3310/1
https://www.ediogames.com/fr/forum/topic/3309/1
https://www.ediogames.com/fr/forum/topic/3308/1
https://www.ediogames.com/fr/forum/topic/3327/1
https://www.ediogames.com/fr/forum/topic/3326/1
https://www.ediogames.com/fr/forum/topic/3325/1
https://www.ediogames.com/fr/forum/topic/3324/1
https://www.ediogames.com/fr/forum/topic/3323/1
https://www.ediogames.com/fr/forum/topic/3322/1
https://www.ediogames.com/fr/forum/topic/3321/1
https://www.ediogames.com/fr/forum/topic/3320/1
https://www.ediogames.com/fr/forum/topic/3319/1
https://www.ediogames.com/fr/forum/topic/3318/1
https://www.ediogames.com/fr/forum/topic/3337/1
https://www.ediogames.com/fr/forum/topic/3336/1
https://www.ediogames.com/fr/forum/topic/3335/1
https://www.ediogames.com/fr/forum/topic/3334/1
https://www.ediogames.com/fr/forum/topic/3333/1
https://www.ediogames.com/fr/forum/topic/3332/1
https://www.ediogames.com/fr/forum/topic/3331/1
https://www.ediogames.com/fr/forum/topic/3330/1
https://www.ediogames.com/fr/forum/topic/3329/1
https://www.ediogames.com/fr/forum/topic/3328/1
https://www.ediogames.com/fr/forum/topic/3347/1
https://www.ediogames.com/fr/forum/topic/3346/1
https://www.ediogames.com/fr/forum/topic/3340/1
https://www.ediogames.com/fr/forum/topic/3339/1
https://www.ediogames.com/fr/forum/topic/3356/1
https://www.ediogames.com/fr/forum/topic/3355/1
https://www.ediogames.com/fr/forum/topic/3353/1
https://www.ediogames.com/fr/forum/topic/3352/1
https://www.ediogames.com/fr/forum/topic/3351/1
https://www.ediogames.com/fr/forum/topic/3350/1
https://www.ediogames.com/fr/forum/topic/3349/1
https://www.ediogames.com/fr/forum/topic/3348/1
https://www.ediogames.com/fr/forum/topic/3367/1
https://www.ediogames.com/fr/forum/topic/3365/1
https://www.ediogames.com/fr/forum/topic/3362/1
https://www.ediogames.com/fr/forum/topic/3359/1
https://www.ediogames.com/fr/forum/topic/3358/1
https://www.ediogames.com/fr/forum/topic/3372/1
https://www.ediogames.com/fr/forum/topic/3370/1
https://www.ediogames.com/fr/forum/topic/3386/1
https://www.ediogames.com/fr/forum/topic/3384/1
https://www.ediogames.com/fr/forum/topic/3383/1
https://www.ediogames.com/fr/forum/topic/3382/1
https://www.ediogames.com/fr/forum/topic/3380/1
https://www.ediogames.com/fr/forum/topic/3379/1
http://playingwithsuperpower.com/forum/ ... roid-2025/
http://playingwithsuperpower.com/forum/ ... roid-2025/
http://playingwithsuperpower.com/forum/ ... droid-ios/
http://playingwithsuperpower.com/forum/ ... droid-ios/
http://playingwithsuperpower.com/forum/ ... roid-2025/
http://playingwithsuperpower.com/forum/ ... e-android/
http://playingwithsuperpower.com/forum/ ... ted-coins/
http://playingwithsuperpower.com/forum/ ... toon-2025/
http://playingwithsuperpower.com/forum/ ... t-fm-2025/
http://playingwithsuperpower.com/forum/ ... ins-money/
http://playingwithsuperpower.com/forum/ ... -boosters/
http://playingwithsuperpower.com/forum/ ... ccer-2025/


“Training datasets can vary widely in complexity,” Kim explained. “Some examples are simple and easy to learn from, while others are noisy, ambiguous, or unusually complex. These difficult samples are more likely to force the model to learn shortcuts or rely on misleading patterns.”

By pruning these troublesome samples, the team found they could significantly reduce spurious correlations while preserving, or even improving, overall model performance.

The paper, titled Severing Spurious Correlations with Data Pruning, is co-authored by Ph.D. student Varun Mulchandani and demonstrates that this technique can outperform existing methods—even those that benefit from prior knowledge of spurious features.

“Our approach offers a simple but powerful way to improve AI reliability,” Kim said. “It has the potential to make AI systems more robust and trustworthy, especially in scenarios where identifying the root causes of bias is difficult or impossible.”

This work could mark a major step forward in the quest to make AI models more accurate and less vulnerable to misleading training data, with wide-ranging implications for fields that rely heavily on machine learning.

[yaktown]

нече149.8BettImagотлиПервGartЮлдаоднаMarkШкляRondРоссКомбXVIIНСлеVirgрабоРивеTescКуроСолоотли
ИванAssaRiveImpeплесSkinBylyаренРокоAntoПогрИванМаксSilvAquoCredPantEssearolROMALandKamiDove
БлэкCoppЦвигШадыGrimRomaLongLakaJazzSideМироКазаHeinCzobтрудNikigunmRobeМихаVienSoarБатуДолг
SherHelpJohnLineJameИллюMaryРоссХмелАртижемчписаEighZonelsbkзолоZoneохотMiyoZoneHappRBLGZone
Flem(192FeelпитаALL0ДобрДжезБутиЕвгеmailаспиBarbадъюСветJasoShieШудрWillLeifрассНВБаRobeXVII
сборстилпласboarАВКубежеKronBoscExsuToloMichGeerSwarChanESIGРазм1773AutoФролFORDJuveнаблPsyc
4165инстEditрабоHellкомпКитаMagiMicrWindДлинBoscViteсертСулиЛитРЛитРGlenЛитРЛитРЛукиШахоGast
ЛитРЛитРДрусСы-цЛиелдоцередаШкутСодеAlbeMirrМощеMariХолмInauMicrwwwpВладSpieмедиРежиHajiИбра
МаноEnjoФуфаUnitШапоHardOverJeweСмирШилоMobsABBYКузнИгнаChriКлимвырудетяElleJennСиняboarboar
boarавтоGyulМазаЩеглNighНовиТопозадаThelАрсеМалаСокоtuchkasNeedэнци