Research Experience

Estimating Sample Complexity in Deep Neural Networks

The most common and effective strategy in deep learning to improve performance and reduce generalization error is to collect more training data \textit{but} the fundamental question: “how much data do we need to achieve certain desirable performance?” still remains open. We aim to present a practical and predictive probabilistic method to model the relationship between training data sizes and misclassification rate.
Paper Github

VisualBackProp for learning using privileged information with CNNs

Proposed a novel training paradigm for CNNs based on learning using privileged information (LUPI) previously proposed for SVMs. Priviledge information (segmentation mask) was incorporated into training by additional loss functions to focus model’s attention on the desired region of the input image. Proposed method was verified through experiments on ImageNet and PASCAL VOC data sets and performance improvements of 2.4% and 2.7% over standard single-supervision model were obtained.
Paper Github

Towards Automated Melanoma Detection with Deep Learning: Data Purification and Augmentation

Proposed a data preparation strategy to tackle detrimental effects of biased data-sets on deep learning models. We utilized partial convolutions for removal of occlusions and synthetic data generation through coupled GAN to balance disease classes. We demonstrate its effectiveness on ISIC 2017 and 2018 grand challenge.
Paper Github

High Frequency Ultrasound Image Segmentation and Analysis

Trained an Active Shape Model to segment brain ventricles of a mouse embryo from its high frequency 3D Ultrasound image. The shape of the brain ventricle was described using a shape context descriptor while principle component analysis was used to generate the model.
Github