Boyang Lyu Email and Phone Number

• Developed Domain Generalization (DG) algorithms via PyTorch for improving the robustness of deep computer vision models; published 7 papers in peer-reviewed journals and conferences.• Implemented the proposed DG algorithms using ResNet-50 as backbone, improving performance by 2% points on average over state-of-the-art methods on 5 DG datasets.• Proposed a novel model selection method for DG, outperforming training-domain validation method across 12 DG algorithms on 3 benchmark DG datasets in selecting more generalizable models.• Designed a machine learning workflow for Brain-Computer Interface (BCI) data analysis, integrating data visualization, anomaly detection, missing value imputation, feature extraction, classification, and evaluation steps using libraries like Pandas, Scikit-learn and NumPy• Implemented multiple time-series classification models like MLSTM-FCN, Time Series Transformer on EEG, fNIRS, and eye-gaze data for human behavior prediction, enhancing the prediction accuracy by 2% compared to baseline methods.• Innovated Gromov-Wasserstein based Domain Adaptation algorithm for mental workload prediction using fNIRS data, increasing accuracy by 5% compared to conventional CNN and RNN models.• Investigated and implemented various anomaly detection algorithms on fNIRS data for data cleaning, boosting the task performance by 3% points.• Collaborated with three teams from different departments on BCI data acquisition, data pre-preprocessing and storage.

• Re-formulated sonic-based cement condition prediction problem from classification to ordinal regression, reducing the MAE loss by 20%.• Architected and implemented machine learning pipelines with a user-friendly interface by shell scripts and Python on Kubernetes.• Designed Domain Generalization algorithms with TensorFlow for improving the robustness of neural networks, achieving a 5% improvement in prediction accuracy over the baseline methods.• Replaced CNN-based backbone with ViT for sonic feature extraction, improving the downstream prediction accuracy by 4%.• Deployed Git for code version control and Docker containers to ensure reproducible environments, streamlining collaboration and enabling efficient development.

• Designed ML workflow for cement condition evaluation utilizing processed sonic data.• Integrated Relational Regularized Auto-Encoder and Variational Auto-Encoder with clustering algorithms like K-means, OPTICS and Agglomerative for automated sonic dataset labeling.• Verified the effectiveness of sonic data in partially capturing cement bond conditions and formation variations.

• Developed string conversion model for chatbot with 95% conversion accuracy.• Tested several NLP models under different application scenarios.

• Designed a thread-based pH sensing bandage, and integrated it with a portable micro-controller for efficient data readout.• Verified the sensitivity of the bandage across various pH environments, and analyzed results using MATLAB.• Established data transmission pathways between the micro-controller and smartphones using Bluetooth for real-time results visualization.