Research

AugmentationVisionCNNNeural Network

Residual Net for Car Detection with spatial transformation

Research Questions: Can we train a rotation-invariant convolutional neural network by designing a multi-head model? How can we formulate a robust learning scheme by proposing a novel augmentation algorithm that scales to emulate a large set of image degradation and occlusion policies?

RNNNeural NetworkSignal ProcessingECGAnomaly Detection

Classification of ECG signals by dot Residual LSTM Network with data augmentation for anomaly detection

Research Questions: Can we propose a conditional variational autoencoder (CVAE) to generate anomalous ECG signals as an augmentation tool? To minimize the number of parameters in the network, multiple dot layers are designed. These dot layers can be considered as non-learnable similarity gates, which can mix features from two branches in the network.