Allan H. Ma

LLM Fine-Tuning for Fraud Detection

Fine-Tuned LLMs on a tabular data set of credit card transactions for fraud detection, and compared its ROC_AUC performance with traditional method, e.g., logistic regression.

This project is done at Scotiabank during Vector Insititute LLM Fine-Tuning and Alignment BootCamp.

Makeup Preview

Generate realistic makeup preview, e.g., lipstick, with unsupervisely learned image-to-image translator based on conditional GANs in Pytorch.

This project is owned by ModiFace Inc.

GAN eval

Experimented with evaluating generated sample qualities based on some divergence metrics across different hyper parameter dimensions.

Paper Github

Generative Adversarial Parallelization

Experimented with parallelized training of multiple Generative Adversarial Networks for improved mode coverage and regularization.

Paper Github

Theano-MPI

Implemented distributed deep learning on ImageNet classification aiming to scale up the training of deep learning models based on data parallelism. It utilizes multiple GPUs on a computing cluster to speed up the training performance.

Paper Github

Software design for oxygen monitoring application

This project aims to build a program for the oxygen monitoring system. The program running on the prototype board (FPGA and MCU) collects oxygen absorption signal and calculates real time oxygen concentration. The prototype includes a large LCD and other human interfaces for signal display, menu control and data recording purposes.