Parameter Reduction Improves Vision Transformers: A Comparative Study of Sharing and Width Reduction

A recent dissertation has addressed the challenges of zero-shot anomaly classification and segmentation, which are essential for detecting anomalies without prior training data. The study formalizes the issue of consistent anomalies, which can bias distance-based detection methods, and introduces CoDeGraph, a framework designed to filter these anomalies effectively.

A new framework named LightHCG has been introduced for glaucoma detection, leveraging HSIC disentanglement and advanced AI models like Vision Transformers and VGG16. This model aims to enhance the accuracy of glaucoma diagnosis by analyzing retinal images, addressing the limitations of traditional diagnostic methods that rely heavily on subjective assessments and manual measurements.

The introduction of PRISM (PRIors from diverse Source Models) marks a significant advancement in dataset distillation, addressing the limitations of existing methods that often rely on a single teacher model. By decoupling architectural priors during the synthesis process, PRISM enhances the generation of synthetic data, leading to improved intra-class diversity and generalization, particularly on the ImageNet-1K dataset.

A comparative analysis was conducted on three image-based methods—VGG16, ViT-B/16, and CoAtNet-Tiny—to classify mental health conditions such as depression and schizophrenia using actigraphy-derived images. The study utilized wrist-worn activity signals from the Psykose and Depresjon datasets, converting them into images for evaluation. CoAtNet-Tiny emerged as the most reliable method, achieving the highest average accuracy and stability across different data folds.

A new method for image clustering, named Hierarchical Semantic Alignment (CAE), has been proposed to enhance the categorization of images by addressing the ambiguity of nouns in semantic representations. This approach integrates caption-level descriptions and noun-level concepts to construct a semantic space that aligns with image features, improving clustering performance without the need for training.