git projects | converging machine

algorithms

Cross-D Conv

This operation introduces a method that enables seamless weight transfer between 2D and 3D convolution operations through Fourier domain phase shifting. This solution addresses the common challenge in biomedical imaging where 3D data, while more applicable, is scarce compared to abundant 2D samples.

Policy Gradient-Driven Noise Mask

A novel deep learning approach that employs reinforcement learning to generate adaptive noise masks for medical image classification. This innovative pretraining pipeline combines a lightweight policy network with a classifier network to tackle the challenges of heterogeneous multi-modal and multi-organ biomedical datasets.

Variational Contrastive Learning

A novel self-supervised learning framework that combines variational methods with contrastive learning to tackle complex classification pretraining tasks. This innovative approach leverages beta-divergence and variational principles to create robust semantic representations, particularly effective for face understanding tasks.

datasets

Refined RadImageNet

A sophisticated toolkit for processing and enhancing the RadImageNet medical imaging dataset, featuring advanced data organization and model benchmarking capabilities. This conversion tool transforms raw radiological images into a refined and stratified structure, optimized for deep learning applications.

YFCC392K Face Attributes Dataset

A comprehensive face attributes dataset containing 392,000 diverse facial images curated from YFCC100M, specifically designed for advancing self-supervised and weakly-supervised face recognition research. This carefully processed collection features weakly-labeled facial attributes, enabling robust model training and evaluation.