Recent news Recent news

Return to Full Page
Back

Two new Google Faculty Research Awards granted to Gergely Neu and Xavier Serra

Two new Google Faculty Research Awards granted to Gergely Neu and Xavier Serra

Two new Google Faculty Research Awards have been granted to DTIC researchers: Gergely Neu and Xavier Serra.

18.03.2019

 

Two new Google Faculty Research Awards have been granted to DTIC researchers: Gergely Neu and Xavier Serra.

UPF press release (catalan, english and spanish)

PI: Gergely Neu

Reinforcement learning (RL) has recently started to receive spotlight as one of the most promising machine learning principles, due to the spectacular results achieved by RL algorithms in solving challenging problems like human-level play in Atari games and the game of Go. Despite these successes, progress in the field remains mainly empirical, with little theoretical understanding of the most effective algorithms. Indeed, most of the deep RL methods used in practice are direct adaptations of classical RL algorithms without directly accounting for the approximation capabilities of the underlying deep networks. The present project aims to address this issue by taking a closer look at the effects of approximations in reinforcement learning methods, leading to a development of new algorithmic tools and a deeper theoretical understanding of existing methods.

 

PI: Xavier Serra, with the participation of Eduardo Fonseca

Large and varied open datasets are essential to advance research in sound event recognition (SER). Google’s AudioSet is very large, but uses YouTube content with usage restrictions. Complementarily, the work-in-progress FSD also uses the AudioSet Ontology and is freely-distributable. This opens the door to organizing public competitions in order to foster research in SER. However, manually labeling massive amounts of data is very time-consuming and expensive. In this project extension, we propose a sustainable ecosystem to foster open research in SER. In this ecosystem, the usage of machine learning approaches (for automatic segmentation and active learning) allows us to curate FSD in a more scalable manner by optimizing human resources. The resulting FSD can be used to organize open machine learning competitions, thus generating knowledge to the community that also feeds back into the dataset creation process.

Full list of awardees:

https://ai.google/research/outreach/faculty-research-awards/recipients/

Multimèdia

Multimedia

Multimedia

Categories: