YongKai DENG

Mathematics A*, Further Mathematics A, Physics A,

1:1 degree

expected 2:1 class degree or above

• Addressed the challenge of sensor overexposure/underexposure during tunnel transitions in autonomous vehicles, leading to object detection failures and safety risks. Developed a novel multi-modal feature fusion approach based on uncertainty estimation to combine data from RGB cameras, LiDAR, and radar for vehicle detection
• Conducted data collection, purification, and simulation using CARLA for scenarios with adverse lighting, occlusion, and weather. Established the theoretical framework and implemented the proposed method, incorporating image uncertainty for adaptive feature fusion. Employed CNN as the baseline model and performed comparative experiments using GNN and RNN architectures. Optimized the network structure, resulting in enhanced safety for autonomous driving in diverse environments.

• Developed a robust anomaly detection system designed to identify defects such as wear, cracks, and other issues in cutting wheels. This involved collecting and preprocessing an extensive dataset of cutting wheel images, including tasks like cleaning, annotation, and augmentation. Conducted thorough research on industrial anomaly detection techniques and implemented various methods including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Autoencoders, and Transfer Learning.
• Adapting pre-trained models to the anomaly detection task, significantly improved performance despite the limited availability of labeled defective data. Throughout the project, collaborated closely with cross-functional teams to understand business requirements, data collection processes, and deployment constraints, ensuring that the solution was both practical and effective.

Developed a neural network model to evaluate the rationality of news headlines, focusing on addressing imbalanced data in text classification tasks. Implemented data enhancement techniques including thesaurus-based substitution, sentence paraphrasing, and kernel density estimation (KDE), integrating these with a Bi-LSTM model and Label Distribution Smoothing (LDS) for improved robustness and accuracy. Utilized the Parrot framework for whole sentence paraphrasing, significantly enhancing data distribution and prediction accuracy. Published a paper comparing various data augmentation methods for text classification at the 2022 IEEE 14th ICCRD as a co-first author.

Collaborated with two students to design and develop a smart Android application using IoT technology and advanced deep learning models. The application collects user activity data via wearable sensors for real-time activity recognition and historical data tracking. Implemented a deep learning model integrating Long Short-Term Memory (LSTM) networks with Fully Convolutional Networks (FCN) for multivariate time series classification. Enhanced model accuracy with a squeeze-and-excitation block, achieving a 99% accuracy rate on testing data.

Developed a small AI mental health assistance chatbot to provide initial psychological support and advice using natural language processing. Utilized Python, the Transformers library, and PyTorch framework to fine-tune the T5 model for mental health counseling. Trained the model on a publicly available dataset from Hugging Face, containing mental health-related questions and positive responses.

Conducted a study on image captioning, evaluating the performance of three decoders (LSTM, Attention LSTM, Transformer) with different encoders (ResNet-18, ResNet-50, ResNet-101) under limited training data conditions. Findings indicated that Attention LSTM excelled with limited data, LSTM was effective with simpler encoders, and the Transformer showed potential with complex encoders and more data. These results offer valuable insights for model selection, suggesting the use of LSTM/Attention LSTM for limited data scenarios with short time dependencies, such as stock market predictions.

Implemented a Transformer model with multi-head attention to improve a neural machine translation (NMT) system for translating German to English. Using PyTorch, developed the multi-head attention mechanism from scratch, allowing the model to focus on different parts of the input sentence simultaneously. The project involved implementing the baseline Transformer architecture and extensive training and evaluation to achieve optimal performance.