Research

Masked autoencoders can overfit low-frequency background structure and underrepresent the high-frequency details needed for recognition. Adds a mutual-information regularizer that aligns encoder representations with high-frequency content without changing the architecture.

Paper

OOD input filtering may not require the cost of a deep encoder when invalid inputs can be described by simple visual cues. Builds a lightweight filter from brightness, contrast, texture, shape, and spatial features with ExtraTrees, then compares it with ResNet-18.

Paper

Semi-supervised segmentation with sparse pixel labels is vulnerable to noisy pseudo-masks and class imbalance. Trains two networks with different loss weightings, uses the stronger prediction as the pseudo-label, and limits consistency learning to saliency-based reliable regions.

Paper Code

The relevant signal occupies a small region of a large image, and fine-grained labels can differ between annotators. Crops the region of interest before classification, ensembles heterogeneous backbones, and masks disputed labels during training.

Paper Code

OOD detectors usually need examples from outside the training distribution, which may be unavailable. Uses CLIPseg to locate the in-distribution object and latent diffusion to remove it, turning training images into synthetic OOD samples.

Paper Code

A classifier can make the right prediction without showing which low-level attributes informed it. Predicts the attributes first, derives the higher-level category from them, and measures each attribute's counterfactual contribution.

Paper

Changes in contrast and image format between training and deployment can alter model behavior. Tests a model trained at one contrast level across controlled combinations of contrast and format to separate their effects.

Paper

Segmentation and classification had to be learned from limited, imbalanced labels in a modality without public data. Combines deep ensembles and TTA with reliable pseudo-labeling that gradually admits high-confidence predictions.

Paper Code

The data vary across institutions, positives are scarce, and lesion locations are unavailable during training. Detects the region of interest before classification and uses class activation maps to localize evidence from image-level labels.

Paper

Random forests struggle to increase each tree's strength without also increasing correlation between trees. Samples triplets from tree split distributions so same-class features move closer and different-class features move apart.

Paper