Chun-Nan Chou

Transfer representation learning for medical image analysis

There are two major challenges to overcome when developing a classifier to perform automatic disease diagnosis. First, the amount of labeled medical data is typically very limited, and a classifier cannot be effectively trained to attain high disease-detection accuracy. Second, medical domain knowledge is required to identify representative features in data for detecting a target disease. Most computer scientists and statisticians do not have such domain knowledge. In this work, we show that employing transfer learning can remedy both problems. We use Otitis Media (OM) to conduct our case study. Instead of using domain knowledge to extract features from labeled OM images, we construct features based on a dataset entirely OM-irrelevant. More specifically, we first learn a codebook in an unsupervised way from 15 million images collected from ImageNet. The codebook gives us what the encoders consider being the fundamental elements of those 15 million images. We then encode OM images using the codebook and obtain a weighting vector for each OM image. Using the resulting weighting vectors as the feature vectors of the OM images, we employ a traditional supervised learning algorithm to train an OM classifier. The achieved detection accuracy is 88.5% (89.63% in sensitivity and 86.9% in specificity), markedly higher than all previous attempts, which relied on domain experts to help extract features.

CLKN: Cascaded Lucas-Kanade Networks for Image Alignment

This paper proposes a data-driven approach for image alignment. Our main contribution is a novel network architecture that combines the strengths of convolutional neural networks (CNNs) and the Lucas-Kanade algorithm. The main component of this architecture is a Lucas-Kanade layer that performs the inverse compositional algorithm on convolutional feature maps. To train our network, we develop a cascaded feature learning method that incorporates the coarse-to-fine strategy into the training process. This method learns a pyramid representation of convolutional features in a cascaded manner and yields a cascaded network that performs coarse-to-fine alignment on the feature pyramids. We apply our model to the task of homography estimation, and perform training and evaluation on a large labeled dataset generated from the MS-COCO dataset. Experimental results show that the proposed approach significantly outperforms the other methods.

Distributed Training Large-Scale Deep Architectures

Scale of data and scale of computation infrastructures together enable the current deep learning renaissance. However, training large-scale deep architectures demands both algorithmic improvement and careful system configuration. In this paper, we focus on employing the system approach to speed up large-scale training. Via lessons learned from our routine benchmarking effort, we first identify bottlenecks and overheads that hinter data parallelism. We then devise guidelines that help practitioners to configure an effective system and fine-tune parameters to achieve desired speedup. Specifically, we develop a procedure for setting minibatch size and choosing computation algorithms. We also derive lemmas for determining the quantity of key components such as the number of GPUs and parameter servers. Experiments and examples show that these guidelines help effectively speed up large-scale deep learning training.

Artificial Intelligence in XPRIZE DeepQ Tricorder

The DeepQ tricorder device developed by HTC from 2013 to 2016 was entered in the Qualcomm Tricorder XPRIZE competition and awarded the second prize in April 2017. This paper presents DeepQ»s three modules powered by artificial intelligence: symptom checker, optical sense, and vital sense. We depict both their initial design and ongoing enhancements.