readme.md

Visualization

Top 100 images flagged by different SOTA methods. Note that they fail to capture the failure mode of duplicate images with conflicting labels.

Use ctrl + scroll (or use pinch to zoom on mobile) to zoom into images.

Top CIFAR100 images detected by SSFT as easy to forget

Figure 1: Top CIFAR100 images detected by SSFT as easy to forget

Top CIFAR100 images detected by SimiFeat as likely noisy

Figure 2: Top CIFAR100 images detected by SimiFeat as likely noisy

High curvature samples from training set according to Slo-Curves (Garg & Roy, 2023)

Figure 3: High curvature samples from training set according to Slo-Curves (Garg & Roy, 2023). Obtained from ResNet18 trained without weight decay on CIFAR100

Calibration curves: Early stopping using curvature of training set as a criterion yields better-calibrated networks.

Model calibration at lowest validation loss stopping (Epoch 17)

Figure 4: Model calibration at lowest validation loss stopping (Epoch 17).

Model calibration at highest curvature stopping (Epoch 19)

Figure 5: Model calibration at highest curvature stopping (Epoch 19).

Model calibration at end of training (overconfident model)

Figure 6: Model calibration at end of training (overconfident model).

Landscape Visualization

Visualizing the loss landscape and decision boundary during training on a toy dataset.

Higher Quality Images

ImageNet, High Curvature Images

ImageNet, Low Curvature Images

Higher Quality Version of Most Memorized according to FZ scores on CIFAR100

Higher Quality Version of Highest Curvature with Weight Decay on CIFAR100