Research
Efficiently Controlling Multiple Risks with Pareto Testing
⇒ Select model configurations that control multiple risks while minimizing additional free objectives.
⇒ Combine Multi-Objective Optimization with Multiple Hypothesis Testing.
⇒ Focus testing on the Pareto Frontier for improved computational and statistical efficiency.
⇒ Application for reliable adaptive computation in large transformer models.
Semi-Supervised
Sound Source Localization Based on Manifold Regularization
⇒ Semi-supervised, utilizes a small number of labelled data and a large number of unlabelled data
⇒ Manifold-regularized optimization, impose a smoothness constraint on possible solutions with respect to a manifold learned in a data-driven manner.
Semi-Supervised Source Localization on Multiple-Manifolds with Distributed Microphones
⇒ Localization based on ad-hoc networks with several nodes, where each node consists of a compact microphone array​.
⇒ Each node represents a different viewpoint that can be associated with a specific manifold.
⇒ Present a Bayesian paradigm for merging the information of the co-related manifolds.
Source Counting and Separation Based on Simplex Analysis
⇒ A novel data-driven geometric approach for source separation, based on a well-established probabilistic model.
⇒ A new representation of the data in a shape of a simplex.
⇒ The simplex dimension corresponds to the number of speakers, and its vertices represent time-intervals with a single dominating speaker.​​
Global and Local Simplex Representations for Multichannel Source Separation
⇒ Combine two simplex representations - global and local.
⇒ Global representation provides probability of activity in each frame.
⇒ Local representation provides dominance of speakers in each frequency.
⇒ Obtain the full spectral mask by aligning each of the local representations using the global probabilities.​​