Ann Arbor Algorithms

Training Deep Convolutional Models

As both GPUs per system and TFLOPs per GPU grow rapidly, how to efficiently preprocess and stream training data to keep the GPUs busy is becoming an increasingly challenging problem. We developed PicPac, a C++ library to efficiently manage and stream massive amount of training data. PicPac fully utilizes the high IOPS of SSD/NVME to support out-of-core random shuffling and stratified sampling, and implements a plug-in framework of data transformation and augmentation to support various training tasks. PicPac's python API is easy to use and is compatible with Tensorflow, PyTorch, MxNet and Caffe.

Medical Imaging and Lesion Detection

We are experienced in deep-learning with DICOM medical images, both 2D and 3D. We have developed deep-learning models to detect and segment lung cancer, breast cancer, multiple-myeloma and other lesions. Our solutions based on PicPac and have ranked high in multiple competitions. See our demo of carotid artery plaque segmentation and 3D reconstruction.

Example of lung nodule detection.

Content-Based Image Search Engine

We developed KGraph, one of today's fastest libraries for approximate nearest neighbor search (benchmark), and Donkey, a NoSQL feature vector database and toolkit for developing nearest neighbor search engines. Donkey supports KGraph and Locality Sensitive Hashing for indexing and supports HTTP/Restful API.
Leveraging KGraph, Donkey and latest deep-learning models for feature extraction, we have helped our client in UK implement a content-based image search engine that indexes tens of millions of images with a single server.

Next Generation Sequencing

A2Genomics is our cloud platform for high-throughput sequencing data analysis and pattern discovery. Our pipeline efficiently processes massivie NGS datasets, run multiple algorithms including PCA, SVD, DESeq, k-means, SOM and WGCNA, and generates publication quality visualizations.

Collaborative Filtering

We have helped a leading Chinese internet radio app with 70+ million users design and implement a recommendation system that minds user behavior and making online personalized recommendations.

Radio Commercial Search and Discovery

We have helped our client in China develop audio fingerprinting algorithms and implement a system that indexes millions of hours of radio broadcast audio covering 100+ cities. The system provides online search-by-example service and automatically discovers repetitive audio clips for new advertisements monitoring.