Software
K-Beam Minimax: Efficient Optimization for Deep Adversarial Learning
- The K-beam minimax algorithm, proposed in the ICML'18 paper [arxiv], allows faster and more stable training of GANs and other adversarial optimization problems such as domain adaptation, privacy preservation, and robust learning.
- The scripts demonstrate the K-beam minimax algorithm applied to different machine learning problems, some implemented in pure python and some with Tensorflow.
- The source code is hosted on GitHub.
Crowd-ML: Crowdsourced Machine Learning with Privacy
- A library for differentially private machine learning with devices including Android and iOS devices, loosely based on our ICDCS'15 paper [pdf].
- The paper has also been featured in a technology news website Gigaom (link).
- The source code is hosted on GitHub.
GRAM: Geodesic Registration on Anatomical Manifold
- GRAM is a framework for groupwise registration of medical images described in our MedIA'10 paper [pdf].
- This work has earned us the 1st place in MedIA-MICCAI Best Journal Paper Award, 2010.
- The source code is hosted on GitHub.
Minimax Filter: Learning to Preserve Privacy from Inference Attacks
- Minimax Filter can preserve privacy of images, audios, or biometric data by making it difficult for an adversary to infer sensitive or identifying information from those data after filtering.
- The library is base on the AISTATS'15 paper [pdf] and the arxiv report [pdf].
- The source code is hosted on GitHub.