Wittawat Jitkrittum

Recorded tutorial on interpretable comparison of distributions and models at NeurIPS 2019. Video. Slides (part 1, part 2, part 3).

I will give a tutorial (with Dougal Sutherland and Arthur Gretton) at NeurIPS 2019 on Mon Dec 9th 11:15 - 13:15 @ West Hall A. The topic will be “Interpretable Comparison of Distributions and Models”. See the event here.

I am invited to give a talk at Swiss Data Science Center. I will speak on informative features for comparing probabilistic models. Slides here. Talk recording here.

A new preprint titled ABCDP: Approximate Bayesian Computation Meets Differential Privacy. We study the interplay between the ABC similarity threshold $\epsilon_{abc}$ and the privacy loss $\epsilon_{dp}$ in differential privacy. Our results show that increasing $\epsilon_{abc}$ (i.e., fast, inaccurate posterior samples) reduces privacy loss, agreeing with our intuition that inaccurate posterior samples reveal less about the underlying sensitive data.

I am honored to be selected to receive ELLIS PhD Award for my thesis titled “Kernel-Based Distribution Features for Statistical Tests and Bayesian Inference”. I would like to thank Arthur Gretton (my PhD supervisor), everyone I met in my PhD journey, and the selection committee who made this possible.

Two papers accepted to NeurIPS 2019.

I will attend DALI 2019 from 3-6 September 2019 in San Sebastian, Spain. Looking forward to meeting everyone!

The first Machine Learning Research School 2019 has concluded. On behalf of the organizing team, I would like to thank all our valued sponsors, all the speakers, the four organizing partners, members of the organizing team, volunteers, and all the participants. The event was a success because of the joint effort from all the involved parties. Stay tuned for the news on the future iteration of the event.

A new preprint “A Kernel Stein Test for Comparing Latent Variable Models”. The kernel Stein discrepancy (KSD) test proposed in Liu et al., 2016 and Chwialkowski et al., 2016 allows one to test the goodness of fit an unnormalized differentiable density model. However, when the model is a latent-variable model with an intractable marginal density, the KSD test cannot be applied easily. In our new work, we extend the KSD test to allow goodness-of-fit testing of latent-variable models whose marginal densities may be intractable.

I will attend ICML 2019 (9-15 June 2019) and CVPR 2019 (16-20 June 2019). If you would like to meet and discuss research, please feel free to send me a message.

Our paper “Kernel Mean Matching for Content Addressability of GANs” (ICML 2019) is now online on arXiv. The link is here. Code coming soon!

I am a co-organizer of the first Machine Learning Research School (MLRS) 2019 to be held in Bangkok, Thailand on 4-11 August 2019. Application is now open. Deadline: 31 May 2019. All welcome to apply! See also our MLRS facebook page.

Our paper titled “Kernel Mean Matching for Content Addressability of GANs” has been accepted to ICML 2019. This work proposes a procedure which adds “content addressability” to any pre-trained implicit model (including GANs). Given a set of input images, the procedure can generate images (from the chosen implicit model) that are similar to the input images, without retraining the model.

Our short paper titled “Generate Semantically Similar Images with Kernel Mean Matching” has been accepted for oral presentation at Women in Computer Vision Workshop at CVPR 2019. Paper here.

I am a co-organizer of the first Southeast Asia Machine Learning (SEA ML) School to be held in Jakarta, Indonesia on 8-12 July 2019. Application is now open. Deadline: 20 April 2019. No registration fee for students.

New preprint: Witnessing Adversarial Training in Reproducing Kernel Hilbert Spaces. In this work, we augment GAN’s loss function with a new term that provides more information about what the generator learns during training.

I spoke at the Symposium on frontier research in information science and technology at VISTEC, Thailand. Topic: Recent Advances in Kernel Methods for Model Criticism. Slides.

We have released Python code for “Informative Features for Model Comparison”.

New preprint: Large sample analysis of the median heuristic. In this work, we theoretically study the commonly used median heuristic for setting the bandwidth of a Gaussian kernel.

Our NIPS paper “Informative Features for Model Comparison” is now online on Arxiv. Python code.