Projects

For this project, I fine-tune a pre-trained BERT on Indonesian to perform named entity recognition task. Specifically, the goal is to train the model to automatically extract the key elements from unstructured raw address. The pre-trained model was adopted from huggingface transformer. The fine-tuned model achieved high f1 score and is able to accurately extract useful elements from new raw address texts.

Full correlation matrix analysis (FCMA) is a recently development fMRI data analysis method. The goal of FCMA is to compare the whole brain interaction pattern between different cognitive states and to identify a set of regions whose interaction significantly differ across the two states. Such computation is extremely computationally demanding but FCMA used a couple tricks to dramatically speed up this computation process. The goal of this project is to go through the 3 main tricks done by FCMA, and demonstrate how FCMA can be implmented with python step-by-step.

This project aims to investigate the neural mechanism that sustain or switch between attentional states. The first experiment combined machine learning framework with non-parametric statistical testing and multivariate analyses to identify neuromarkers that characterize and differentiate between attentional states. The second experiment built upon these findings by using the pre-trained model and a real-time neuro-feedback framework to design a closed-loop attentional focus traninig framework for potential clinical uses.

We build an API for predicting the popularity of Mandopop/C-pop song. You can simply upload a song, and the model will tell you the estimated popularity of it. The model is based on 1. an inception CNN which takes mel-spectrogram of the song as input and 2. a fully-connected DNN which takes auto-tagging information as input.

This project primarily focused on 2 hot problems in fMRI research. First, we provided insights on how to significantly increase effect size for fMRI research such that even with relatively small sample size, we would still have enough statistical power to detect such effects. Second, we explore an useful heuristic that guide feature selection to improve predictive models for human cognitive functions.