Blaine Hoak
@blainehoak.bsky.social
Ph.D. candidate in Computer Sciences @UWMadison advised by @pdmcdan | secure & trustworthy AI/ML | prev. Ph.D. Intern @Visa Research | she/her 🏳️🌈
Reposted by Blaine Hoak
🪵🧵 Texture vs. shape. Blaine Hoak dives into real-world evidence of texture bias in vision models. Paper: arxiv.org/abs/2412.10597 #SaTML25
April 10, 2025 at 3:20 PM
🪵🧵 Texture vs. shape. Blaine Hoak dives into real-world evidence of texture bias in vision models. Paper: arxiv.org/abs/2412.10597 #SaTML25
Interested in reading more?
📄 Paper: arxiv.org/abs/2412.10597
💻 Code: github.com/blainehoak/e...
⭐️ Project Page: hoak.me/publication/...
📄 Paper: arxiv.org/abs/2412.10597
💻 Code: github.com/blainehoak/e...
⭐️ Project Page: hoak.me/publication/...
Err on the Side of Texture: Texture Bias on Real Data
Bias significantly undermines both the accuracy and trustworthiness of machine learning models. To date, one of the strongest biases observed in image classification models is texture bias-where model...
arxiv.org
June 20, 2025 at 6:00 PM
Interested in reading more?
📄 Paper: arxiv.org/abs/2412.10597
💻 Code: github.com/blainehoak/e...
⭐️ Project Page: hoak.me/publication/...
📄 Paper: arxiv.org/abs/2412.10597
💻 Code: github.com/blainehoak/e...
⭐️ Project Page: hoak.me/publication/...
Finally, natural adversarial examples—samples that are naturally occurring but confidently misclassified—are also a consequence of texture bias: up to 90% of these samples contain misaligned textures, explaining why they are confidently mispredicted. Pic: nat adv examples classified as honeycombs
June 20, 2025 at 6:00 PM
Finally, natural adversarial examples—samples that are naturally occurring but confidently misclassified—are also a consequence of texture bias: up to 90% of these samples contain misaligned textures, explaining why they are confidently mispredicted. Pic: nat adv examples classified as honeycombs
We find that models rely heavily on specific textures learned during training in order to accurately and confidently predict images. When these textures are not present, model accuracy can decrease by as much as 66 percentage points!
June 20, 2025 at 6:00 PM
We find that models rely heavily on specific textures learned during training in order to accurately and confidently predict images. When these textures are not present, model accuracy can decrease by as much as 66 percentage points!
In this work, we find that texture information plays a critical role in real-data decisions and model trustworthiness. With new metrics, we identify textures in natural images (images with grid texture shown below) and analyze how model behavior changes in the presence of different textures.
June 20, 2025 at 6:00 PM
In this work, we find that texture information plays a critical role in real-data decisions and model trustworthiness. With new metrics, we identify textures in natural images (images with grid texture shown below) and analyze how model behavior changes in the presence of different textures.