Projects

DiffMusic: A Zero-shot Diffusion-Based Framework for Music Inverse Problem

Jia-Wei Liao, Pin-Chi Pan, and Sheng-Ping Yang, Deep Learning for Music Analysis and Generation Final Project, 2024.

Slides

Code

Taiwan-LLM Tutor: Revolutionizing Taiwanese Secondary Education with Large Language Model

Jia-Wei Liao, Ji-Jia Wu, Kun-Hsiang Lin, and Kang-Yang Huang, Applied Deep Learning Final Project, 2023.

Report

Slides

Code

Multimodal Pathological Voice Classification

Chun-Hsien Chen, Shu-Cheng Zheng, Jia-Wei Liao, and Yi-Cheng Hung, AICUP, 2023

Report

Code

Optimizing Marketing Strategies and Data Analysis for Bag Brand

Yi-Jhen Ciou, I-Ting Wang, Kai Yu, and Jia-Wei Liao, TMBA Final Project, 2023

A Small Object Detection Framework for Unmanned Aerial Vehicles Images

Jing-En Huang and Jia-Wei Liao, AI CUP, 2022.

Report

We developed a state-of-the-art one-stage detection model as the baseline framework for the competition, with the goal of accurately detecting small objects. As a result, we achieved a top 5 ranking in the competition with a public score of 0.7394 and a private score of 0.7550.

Crop Image Recognition

Jia-Wei Liao, Yen-Jia Chen, Yi-Cheng Hung, Jing-En Hung, and Shang-Yen Lee, AI CUP, 2022.

Report

Code

We trained EfficientNet and Swin Transformer model, incorporating auto-augmentation techniques to enhance dataset diversity. To assess the stability and benefits of ensemble methods, we applied Test-Time Augmentation (TTA) and ensemble strategies to both public and private datasets. This approach resulted in a public ranking of 9th and a private ranking of 8th, with scores of 0.9329 and 0.9344, respectively.

Explainable Information Tagging for Natural Language Understanding

Jia-Wei Liao, Jung-Mei Chu, and Chia-Chi Huang, AI CUP, 2022.

Report

Code

We constructed an explainable deep learning model for a NLP task by converting it into a summary task. We applied the T5 model to two datasets with different pre-processing methods and surpassed the original data's baseline. We achieved a public score of 0.801772 and a private score of 0.85190.

Contour Segmentation for Spread Through Air Spaces in Lung Adenocarcinoma

Jia-Wei Liao, Kuok-Tong Ng, and Yi-Cheng Hung, AI CUP, 2022.

Report

Code

We developed a UNet-based model with a diverse backbone, addressing the imbalance of foreground and background through a combination of DiceLoss and Focal Loss, and enhancing robustness through auto-augmentation. Our post-processing strategies improved accuracy, resulting in a 3rd place public ranking and 16th place private ranking, with scores of 0.9194 and 0.9109 respectively.

Supervised Learning for Few-Shot Orchid types Classification with Prior Guided Feature

Yu-Hsi Chen, Jia-Wei Liao, and Kuok-Tong Ng, AI CUP, 2022.

Report

Code

We developed a few-shot learning approach to accurately distinguish visually similar orchid species, addressing challenges posed by biotechnology-created species and limited training data. Using various models, data augmentations, and optimization techniques, we ensembled top predictions, achieving a Macro-F1 score of 0.9115 (public) and 0.8096 (private), ranking 15th out of 743 teams.

MediaTek Low-power Segmentation Competition

Jia-Wei Liao, Machine Learning Final Project, 2022.

Code

We proposed a lightweight semantic segmentation model optimized for embedded systems and traffic scene analysis in Asian regions, including Taiwan. The model balances accuracy, low power consumption, and real-time performance, designed for MediaTek's Dimensity Series platform. Evaluated using mIoU, a standard metric for semantic segmentation, it achieved a public test score of 0.5624.

Ultrasound Nerve Segmentation

Jia-Wei Liao, Kuok-Tong Ng, and Yi-Cheng Hung, VRDL Final Project (Kaggle), 2021

Report

Code

We propose a deep learning method for ultrasound nerve segmentation using UNet with an EfficientNet backbone. Two novel techniques, Erosion Mask Smoothing (ELS) and adaptive Single Model Ensemble (ASME), improved performance. Evaluated with the Dice Similarity Coefficient (DSC), our approach achieved a private test score of 0.7234, surpassing the baseline.