Mostafa ElAraby
@mostafaelaraby.bsky.social
AI researcher @ Mila, UdeM. PhD focused on OOD detection & generalization. Building robust deep learning. Previously: Microsoft ATL, Tensorgraph. #AI #MachineLearning
GROOD is a practical, post-hoc framework for robust OOD detection & safer AI.
Huge thanks to the team: Mostafa ElAraby, Sabyasachi Sahoo, Yann Pequignot, Paul Novello, Liam Paull.
And collaborators: @Mila_Quebec, @DIRO_UdeM, @UniversiteLaval, & the DEEL project.
Huge thanks to the team: Mostafa ElAraby, Sabyasachi Sahoo, Yann Pequignot, Paul Novello, Liam Paull.
And collaborators: @Mila_Quebec, @DIRO_UdeM, @UniversiteLaval, & the DEEL project.
September 25, 2025 at 11:20 PM
GROOD is a practical, post-hoc framework for robust OOD detection & safer AI.
Huge thanks to the team: Mostafa ElAraby, Sabyasachi Sahoo, Yann Pequignot, Paul Novello, Liam Paull.
And collaborators: @Mila_Quebec, @DIRO_UdeM, @UniversiteLaval, & the DEEL project.
Huge thanks to the team: Mostafa ElAraby, Sabyasachi Sahoo, Yann Pequignot, Paul Novello, Liam Paull.
And collaborators: @Mila_Quebec, @DIRO_UdeM, @UniversiteLaval, & the DEEL project.
The results speak for themselves! 🏆
GROOD achieves SOTA performance on major benchmarks. Highlights:
🔹 84.44% Far-OOD AUROC on CIFAR-100
🔹 94.8% Far-OOD AUROC on ImageNet-1K
🔹 Robust on Transformers (ViT & Swin-T) where others fail!
GROOD achieves SOTA performance on major benchmarks. Highlights:
🔹 84.44% Far-OOD AUROC on CIFAR-100
🔹 94.8% Far-OOD AUROC on ImageNet-1K
🔹 Robust on Transformers (ViT & Swin-T) where others fail!
September 25, 2025 at 11:20 PM
The results speak for themselves! 🏆
GROOD achieves SOTA performance on major benchmarks. Highlights:
🔹 84.44% Far-OOD AUROC on CIFAR-100
🔹 94.8% Far-OOD AUROC on ImageNet-1K
🔹 Robust on Transformers (ViT & Swin-T) where others fail!
GROOD achieves SOTA performance on major benchmarks. Highlights:
🔹 84.44% Far-OOD AUROC on CIFAR-100
🔹 94.8% Far-OOD AUROC on ImageNet-1K
🔹 Robust on Transformers (ViT & Swin-T) where others fail!
A picture is worth a thousand words. 📸
While the standard feature space (left) is messy, our gradient-aware space (right) creates a much clearer separation between In-Distribution (blue) and OOD (red/orange) samples!
#DataScience #ComputerVision
While the standard feature space (left) is messy, our gradient-aware space (right) creates a much clearer separation between In-Distribution (blue) and OOD (red/orange) samples!
#DataScience #ComputerVision
September 25, 2025 at 11:20 PM
A picture is worth a thousand words. 📸
While the standard feature space (left) is messy, our gradient-aware space (right) creates a much clearer separation between In-Distribution (blue) and OOD (red/orange) samples!
#DataScience #ComputerVision
While the standard feature space (left) is messy, our gradient-aware space (right) creates a much clearer separation between In-Distribution (blue) and OOD (red/orange) samples!
#DataScience #ComputerVision
What's the secret sauce? ✨
Instead of just feature distance, GROOD uses a synthetic OOD prototype to measure a sample's gradient sensitivity.
ID samples are stable, while OOD samples are sensitive. This creates a powerful, clear signal for detection!
Instead of just feature distance, GROOD uses a synthetic OOD prototype to measure a sample's gradient sensitivity.
ID samples are stable, while OOD samples are sensitive. This creates a powerful, clear signal for detection!
September 25, 2025 at 11:20 PM
What's the secret sauce? ✨
Instead of just feature distance, GROOD uses a synthetic OOD prototype to measure a sample's gradient sensitivity.
ID samples are stable, while OOD samples are sensitive. This creates a powerful, clear signal for detection!
Instead of just feature distance, GROOD uses a synthetic OOD prototype to measure a sample's gradient sensitivity.
ID samples are stable, while OOD samples are sensitive. This creates a powerful, clear signal for detection!