How do smart speakers, phones and connected devices understand what we’re saying?
This month we look at how the interaction between machines and our voices works now and how can we build even more intelligent systems for complex interaction in the future.
With the answers, we have Dr Giovanni di Liberto, who researches human perception and Dr Andrea Patane, whose work is on what happens when machine learning fails, and what we can learn from that.
TOPICS WE DISCUSSED INCLUDE
01:08 How computers and intelligent systems ‘perceive’ what humans are saying
05:37 The processes at work in living brains and artificial machine learning models are each ‘black boxes’ that we need to understand better
08:45 How our understanding of human learning is helping to influence machine learning models?
12:11 How can intelligent systems ‘learn’ or adapt to nuances, such as different accents in speech?
16:40 If intelligent systems ‘misunderstand’ something, it could have disastrous effects in some situations. What ethics and safety issues come into play?
26:39 How can research into machine learning and perception help to improve technology that can help us support senses such as hearing?
30:00 As machine learning drives forward, we need to be mindful of the assumptions on which it is based?
Dr Giovanni Di Liberto is with TCD School of Computer Science and Statistics. Giovanni’s scientific interests centre on understanding the brain mechanisms underlying speech comprehension. In his work, he develops data analysis methods and applies them to brain data to identify the neural processes responsible for the transformation of a sensory stimulus into its abstract meaning. Brain electrical data is measured with either non-invasive (e.g., electroencephalography – EEG) or invasive (e.g., electrocorticography – ECoG) technologies. The first aspect of his research is methodological and has produced novel experimental and analysis frameworks to investigate cortical auditory processing. The second aspect of his research is to use such novel methods to test theories on auditory perception, such as the hierarchical processing of speech and predictive processing theories (e.g. predictive coding). Finally, the third part of his work is translational and involves the identification of solutions to utilise his novel methods in applied settings, for example as tools to develop brain-computer interfaces (COCOHA project) or as objective measures for the monitoring of language development and healthy ageing.
Dr Andrea Patane has been an Assistant Professor at Trinity College Dublin since May 2022. His main area of research is safety and robustness, with a focus on providing formal guarantees for machine learning models learnt from data, possibly with probabilistic components and/or operating in uncertain and adversarial environments. The development of such guarantees able to deal with uncertainty is of paramount importance if intelligent systems are to be deployed in scenarios where the safety of individuals is of concern, such as biomedical applications and autonomous driving.