Chun-Min Chang

Sr. Algorithm Engineer (ML) at Ambarella Corporation, Santa Clara, U.S.A., 2020/06 - Now

Accelerate model inference by optimizing bit width of tensors in post-training quantization and folding operations in a frozen computation graph. Also, implement specialized primitives and arithemtic operations for various target chips with GPU acceleration.

Design and deploy lightweight deep learning models on iOS and Android platforms for photo style transfer, photo super-resolution, low-light photo enhancement, and object instance segmentation.

Apply cutting-edge machine learning techniques to interdisciplinary studies, and publish research works in Nature Digital Medicine, Nature Communications, Computer methods and programs in biomedicine.

We developed a deep learning approach for automatically determining the estimated glomerular filtration rate (eGFR) and chronic kidney disease (CKD) status over 4,505 kidney ultrasound images. Overall CKD status classification accuracy of our model was 85.6% —higher than that of experienced nephrologists (60.3%–80.1%).

Using over 20,000 images with GPS locality information belonging to nearly 2,000 moth species, our approach learned deep representations that accurately predict each species’ mean elevation based on colour and shape, as well as help identify the underlying mechanisms generating this biogeographic pattern.

We proposed a novel training algorithm for convolutional neural networks to achieve dynamic inference at run time and enhance model robustness against adversarial attacks without any extra computational cost or memory overhead.

We propose the very first non-intrusive measurement methodology for quantifying the timing and positioning accuracy of commodity Virtual Reality (VR) systems. Our methodology considers the VR system under test as a black-box and can work with any VR applications without code instrument or system modification.

We leveraged signal processing and machine learning to transform in-the-wild sensor data from the wearable device into intelligent health experiences. Training over 1,200 records from 600 fasting participants, our non-invasive approach showed over 93.5% of test samples in zone A of the Clarke consensus error grid.

We design a unsupervised model for differentiating histology tissue matrices. The learning algorithm is guided by superpixel refinement and deep feature clustering, and beats other unsupervised methods in metrics of accuracy and inference time.