Project: Distribution embedding network for meta-learning with variable-length inputGlassbox Machine Learning Team, Google ResearchHosts: Sen Zhao & Mahdi Milani Fard1. Formalized a meta-learning framework for learning the optimal aggregation rule to combine base classifiers.2. Proposed and implemented a novel meta-learning approach for applications where both the data distribution and the number of features could vary across tasks.3. Developed a novel methodology to massively simulate binary classification training tasks for the proposed approach.4. Demonstrated significant improvement compared to existing methods in numerical studies.

Project: Search by EmbeddingsMusic ML TeamManager & Mentor: Fabian Moerchen & Brandyn Kusenda1. Developed a track-query joint embedding model (TraQue) to find relevant tracks for any queries user may ask.2. Collected, analyzed, and preprocessed data for modeling using PySpark.3. Utilized pre-trained query embeddings (BERT) and track embeddings (Cuttlefish) as building blocks.4. Demonstrated significant improvements compared to the baseline.5. Showed high precision in terms of finding relevant tracks on some examples where the current system has zero or low precision.

Project: Nonparametric K-sample test Adviser: Xuegong Zhang, Department of Automation, Tsinghua University1. Proposed a scalable density divergence based test for K-sample test problem by segmentation.2. Demonstrated significant improvement compared to other tests in detecting heteroscedasticity and nonlinearity.

Project: Bayesian structure learning for stationary time series Adviser: Emily Fox, Ph.D & Nicholas Foti, Ph.D, Department of Statistics, University of Washington1. Designed a complex multi-frequency version of Birth-Death MCMC algorithm (CBDMCMC) to learn graphical structure of high-dimensional stationary time series.2. Proposed a novel sampler for complex G-Wishart distribution by means of blocked Gibbs sampling.3. Implemented the CBDMCMC algorithm in C++ and integrated it in the R package BDgraph.