UNIVERSITY OF HERTFORDSHIRE            
       COMPUTER SCIENCE RESEARCH COLLOQUIUM       
                                                  
                     presents   
					                   
	"Trustworthy AI in Speech and Language Processing: 
	     Reliability, Transparency and Privacy"                
                      				                        
               Prof. Dorothea Kolossa                  
                                                  
         (Ruhr-Universitaet Bochum, Germany)           
                                                  
                10 February 2020                  
                                                  
                  13:00 - 14:00                   
                                                  
                   Over Zoom          
		 
Abstract

Machine learning and AI technology for speech and audio has expanded to
encompass a multitude of new directions. With access to relevant data,
one can build psychological profiles that rival the understanding of
long-term partners, identify speakers and understand speech at the level
of human listeners, and even surpass human capabilities when it comes to
complex tasks, such as multi-modal source separation in the noisiest of
scenarios.
 
However, the prevalent approaches to AI have some properties that render
them problematic in real life. For one, AI models can be overconfident
and unreliable, which is already an issue with out-of-domain data, and
even more so when a system is under attack. This is especially a problem
when its models are treated as a black box, opening the door to explicit
manipulation as well as implicit bias. Finally, the capabilities of AI
are fundamentally dependent on large datasets, the acquisition of which
conflicts directly with human desire for privacy.
 
This talk will therefore discuss three desiderata which I believe to be
fundamental in trustworthy AI:
reliability, transparency and privacy-preservation.
I will discuss technical aspects of each of these issues, with a focus
on reliability in the face of adversarial input, and invite discussion
on these desiderata, and on recent advances to support their realization.