Research
Object Detection
- Developing a semi-supervised object detection method by combining generative adversarial networks (GANs) with spatial transformer networks.
- This project addresses the challenges posed by limited labeled training data in a semi-supervised setting for detecting environmental indicators in Google Street View images.
- Developing an innovative transfer learning model based on the DEtection TRansformer (DETR) architecture for object detection.
- This project aims to leverage pre-trained models and adapt them to new domains to enhance the accuracy and efficiency of object detection systems.
Adversarial Robustness
- Developed a novel classification method that is robust against black-box adversarial attacks without requiring any prior knowledge about the attack.
- This method is achieved by "inverting" a conditional generative network.
- This research addresses the vulnerability of machine learning models to adversarial attacks. The proposed method enhances the security and reliability of the classification process.
Limited Data
- Multi-task Learning: Developed a multi-task classifier that improves performance on a limited amount of labeled data by leveraging additional unlabeled data.
- The multi-task classifier developed in this research improves the accuracy and generalization capabilities for classifying Google Street View images.
- Active search: Developed an active search method for finding valuable items from a large pool of unlabeled data points for regression.
- This method improves the efficiency and accuracy of regression models when dealing with extensive data collections without labeled annotations.
Image Generation and Denoising
- Generation: Developed a multi-resolution network that combines deep neural networks and disjunctive normal shape model (DNSM) to efficiently reconstruct binary images at various resolutions, particularly when high-resolution training images are scarce.
- This research focuses on image generation, and enables efficient reconstruction of binary images at arbitrary resolutions, even in scenarios where high-resolution training images are not readily accessible.
- Denoising: Developed a novel convolutional denoising autoencoder combined with a disjunctive normal shape model (DNSM-AE).
- This approach successfully reconstructs clear and sharp denoised shapes by leveraging their indicator function representation.
Detection of Human Attention
Image Credit: peterschreiber.media/Shutterstock
- Designed and implemented a novel algorithm to record EEG signals using an Emotiv device.
- Implemented signal processing and machine learning methods to classify human attention into two classes, resulting in over 92% accuracy.
- The research has applications in areas such as cognitive neuroscience, mental health monitoring, and attention-based user interfaces, where accurately detecting attention levels is crucial.