- Mid-term report due date: 9 Nov, 2015
- 4 pages, single column
- Final Report due date: 5 Dec, 2015
- 8 pages, single column
- Presentation: 7 Dec, 2015
- 20 minutes
Mini-projects (5 credits)
To study theoretical and experimental results of an existing paper with new results (theoretical or practical). Here new result means some theoretical or experimental result that has not been covered in the paper. This new result will be more appreciated if it tries to explore some limitations of the existing methods which are rarely discussed in the papers.
Full project (8 credits)
To study theoretically and experimentally a novel algorithm which can be extension of some existing paper or can be a novel concept.
Differentially private topic models (8 credits)
Topic models have been established as one key discipline in machine learning over the last decade. Topic models are applied in various applications from text analysis to social media etc. Due to increased concern on privacy, a study of differentially private topic model is timely with no prior work. This project intends to explore that on the basic topic model namely latent Dirichlet Allocation (LDA).
Siddharth and Rakshith
Practical One Posterior Sample Algorithm for Differential Privacy (8 credits)
One posterior sample (OPS) is an elegant concept proposed by Wang et al that says that one sample from the posterior distribution is differentially private. The idea is theoretically supported however that makes it practically hard to apply. This projects intends to study some practical approach to apply OPS is more practical scenario.
Differentially private logistic regression (5 credits)
Logistic regression (LR) is one of the most popular and basic classifier in the history of machine learning. Differentially private logistic regression has been proposed by Chaudhuri et al and this mini-project intends to study that paper theoretically and experimentally.
A Simple and Practical Algorithm
for Differentially Private Data Release (5 credits)
Hardt et al proposed a simple algorithm for differentially private data release, based on a simple combination of the Multiplicative Weights update rule with the Exponential Mechanism to be called as MWEM. MWEM algorithm achieves what are the best known and nearly optimal theoretical guarantees, while at the same time being simple to implement and experimentally more accurate on actual data sets than many existing methods. This mini-project intends to explroe this work theoretically and experimentally.