Research Projects

Back to Home

Relationship of Data Complexity and Model Robustness

Peking University; Advisor: Prof. Zhanxing Zhu; Oct. 2018-present

  • Did a paper review on theoretical analysis about network robustness and generalizability.
  • Experimentally discovered the relationship between robustness and data amount and input dimension.
  • Tried introducing randomness into network weights and feature maps to improve robustness.

Defense of Adversarial Examples vis Sparse Coding [Arxiv]

UCSD; Advisor: Prof. Hao Su; Jul. 2018-Sept.2018

  • Proposed patch-based Introspective Neural Network for restoration of adversarial examples on ImageNet.
  • Proposed an attack-agnostic adversarial defense via convolutional sparse coding and outperformed state-of-the-art by up to 26.6%.
  • Organized and finished an academic paper from scratch.

3D Multi-view Point Cloud Reconstruction from Single Image [AAAI 2019 Oral, pdf]

MSRA; Advisor: Jinglu Wang; Dec. 2017-Mar. 2018

  • Proposed a 2D convolution-favored 3D point cloud representation recording coordinates and visibilities.
  • Proposed a post-processing network to refine conversion from point cloud to voxel achieving about 5% IoU shift.
  • Implemented multi-view point regression network which outperformed state-of-the-art by 0.1 IoU on hard classes.

3D Indoor Scene Reconstruction

MSRA; Advisor: Jinglu Wang; Mar. 2018-May 2018

  • Reimplemented Roomnet, a proprietary architecture for room key-point detection.
  • Proposed multi-channel boundary map denoting neighboring and vertical relationships in room plane reconstruction.

Selected Presentations

Neural Process [slide]

Prof. Hao Su's Lab; UCSD

This pre first reviews Gaussian Processes and Variational Auto-encoder. Then it introduces a graph model family - neural processes.

Network Uncertainty [slide]

Prof. Hao Su's Lab; UCSD

This pre introduces two papers related to network uncertainty. The first is Bayesian approximating through dropout. The second is measuring uncertainty for adversarial examples.

Testing Bayesian Network [slide]

Prof. Adnan Darwiche's Lab; UCLA

This pre summarizes Testing Bayesian Network(TBN) which compile Bayesian network queries to arithmetic circuit. Then it posts some discussion and potential applications of TBN.

Network Visualization and Interpretation [slide]

Prof. Yizhou Wang's Lab; PKU

This pre first classifies network visualization and interpretation. Then it introduces several works for every type of method.

Style Transfer [slide]

Prof. Yizhou Wang's Lab; PKU

This pre introduces concept, basic method and several improvements of style transfer.

Generative Adversarial Networks [slide]

Prof. Yizhou Wang's Lab; PKU

This pre introduces the background, method, applications and discussion of generative adversarial networks(GAN).