Jiaqi Guan

University of Illinois at Urbana-Champaign

Ph.D. Student (2019.08 - Now)

Student Intern (2017.06 - 2017.09)

Department of Computer Science

Tsinghua University

2014.08 – 2018.07

Bachelor of Engineering in Automation

GPA: 89 / 100

Ranking: 13 / 135

Carnegie Mellon University

2018.09 - 2019.02

Robotics Institute, School of Computer Science

Short-term Research Scholar

First-person Activity Forecasting [pdf]

Jiaqi Guan, Ye Yuan, Kris Kitani, Nick Rhinehart

We develop an efficient model that can properly represent the joint space of discrete and continuous behaviors in order to predict a person's future behaviors. We learn a deep generative model across discrete actions and continuous positions by leveraging recent advances in discrete generative modeling and continuous conditional generative modeling. We evaluate our method on a large-scale egocentric dataset, EPIC-KITCHENS, demonstrate the effectiveness of joint trajectory-action forecasting, and observe our method exhibits performance superior to that of related generative models.

Energy-efficient Amortized Inference with Cascaded Deep Classifiers [pdf]

Jiaqi Guan, Yang Liu, Qiang Liu, Jian Peng

Deep neural networks have been remarkable successful in various AI tasks but often cast high computation and energy cost. We address this problem by proposing a novel framework that optimizes the prediction accuracy and energy cost simultaneously. In our framework, each data instance is pushed into a cascade of deep neural networks with increasing sizes, and a selection module is used to sequentially determine when a sufficiently accurate classifier can be used for this data instance. The cascade of neural networks and the selection module are jointly trained in an end-to-end fashion by the REINFORCE algorithm to optimize a trade-off between the computational cost and the predictive accuracy.

Generation of Synthetic Electronic Medical Record Text [pdf]

Jiaqi Guan, Runzhe Li, Sheng Yu, Xuegong Zhang

Free EMR texts are lacking consistent standards, rich of private information, and limited in availability, which hinders the development of ML and NLP methods for EMR data analysis. To tackle these problems, we developed a model to generate synthetic text of EMRs called Medical Text Generative Adversarial Network. It takes disease features as inputs and generates synthetic texts as EMRs for the corresponding diseases. We evaluate the model from micro-level, macro-level and application- level on a Chinese EMR text dataset. The results show that the method has a good capacity to fit real data and can generate realistic and diverse EMR samples.

Accurate People Detection

Jiaqi Guan, Feng Chen

People detection has many practical uses, such as detecting whether there are suspicious people around fences, people counting and so on. We first applied some famous visual detection frameworks such as Faster-RCNN, YOLO, but the performance was not satisfactory. Finally, we reproduced feature pyramid network based on Caffe that had a better performance on small object detection. We also made a simple people counting mobile app to test the algorithm.