Projects

SGM-PINN

By leveraging probabilistic graphical models (PGMs), SGM-PINN allows adaptively forming much smaller epochs by only selecting the most representative data samples based on a graph-based importance sampling strategy. Our preliminary results show that SGM-PINN can achieve up to 50% faster convergence for training PINNs on several computational fluid dynamics (CFD) problems.

SPADE(ICML 2021)

SPADE measures the adversarial robustness of an ML model by examining the bijective distance mappings between the input/output graph-based manifolds. Moreover, SPADE can be further used to reveal the robustness of input data, which guides downstream applications such as adversarial training. Compare with the VANILLA PGD training, SPADE can achieve up to 18% better accuracy.

Hair Detection Robustness Improvement(Industrial Task)

This work manipulates robust training and data augmentations on a tiny dataset(only 200 pictures). It enables hair detection and achieves 99.97% accuracy on the LEMA (Beijing) Technology Co., Ltd model. Moreover, test accuracy barely drops under severity 5 corruption, which means the reasonable perturbation(e.g., brightness difference or blur) will not affect the accuracy.

Vial Classification Accuracy & Runtime Improvement(Industrial Task)

This work exploits data selection, neural network pruning, and data augmentations to accelerate Vial classification while performing a better accuracy, leading to a 42% accuracy improvement and 5.8x runtime speedup over the previous model from LEMA (Beijing) Technology Co., Ltd.

Benchmarking Tools(Industrial Task)

I participated in building benchmarking tools for AI security and robustness. This project included: 1. Creating a multi-level vulnerability assessment to identify inherent model vulnerabilities across data, model, and system layers. 2. Developing a lightweight detection for fine-grained detection of malicious samples, poisoned models, and backdoor access. 3. Implementing a robust enhancement for data and algorithms, providing trustworthy protection, from data collection to decision inference. 4. Constructing a security measurement with verifiable and explainable decisions.

CanSat Design(Industrial Task)

A CanSat is a type of rocket payload used to teach space technology. This work focuses on CanSat PCB design and final packaging for Shanghai ASES Spaceflight Technology Co.Ltd..

S2D: Sample-to-Decision-Boundary Distance

S2D, a data robustness ranking technique, measures sample manifold distance to decision boundaries and predicts clean data robustness after perturbation. S2D significantly improves robustness evaluation with smaller dataset ranges and massively accelerates training without decreasing accuracy. Our experiments show promising empirical results for neural networks trained with the CIFAR-10 and CIFAR-100 data sets.

SAGRAM

SAGRAM is a black-box efficient and effective coreset selection algorithm. By exploiting Gaussian graphical models, SAGRAM significantly improves the training efficiency without compromising the accuracy of deep models.

-->