There were multiple speakers from both academia and industry who covered many different aspects of AI, from language modelling (Michele Catasta from Google showed tools capable to map natural language into code), to computer vision (Georgia Gkioxary from Meta showed how to bridge he semantic gap between pictures and their interpretation), intelligent transportation (Alexandre Alahi from EPFL showed how to make self driving vehicles socially intelligent) and others.
However, despite the diversity and the variety of the topics addressed in the symposium, there were a few points that kept being made by all speakers:
It is time to outline positive visions of AI capable to orient the future developments of the field towards positive impacts for humanity;
Data plays a major role and their availability risks to generate a divide between those who can get the data and those who cannot;
There is a time before the advent of Deep Learning and a time after he advent of Deep Learning, as a matter of fact, AI and Deep Learning are today synonymous;
A great idea about AI or its application cannot have success without working on the engineering aspects related to its deployment.
The organisers promise to make the content available online soon.
G.Rennie, O.Perepelkina and A.Vinciarelli, “Which Model is Best: Comparing Methods and Metrics for Automatic Laughter Detection in a Naturalistic Conversational Dataset“, Proceedings of Interspeech, 2022.
A.Bîrlădeanu, H.Minnis and A.Vinciarelli, “Automatic Detection of Reactive Attachment Disorder Through Turn-Taking Analysis in Clinical Child-Caregiver Sessions“, Proceedings of Interspeech, 2022.
Both works gave me the possibility to explore new topics and to collaborate with new students and colleagues. The first work deals with a problem that was addressed for the last 10 years at least, but it is still challenging because of the lack of realistic data. Not surprisingly, the key-point of the work is exactly that we experiment on one of the biggest databases ever (several hundreds of spontaneous laughter events). The. second work deals with an elusive and neglected child psychiatry pathology, i.e., the Reactive Attachment Disorder (RAD). It turns out to be another approach based in the analysis of the turn-taking (the way people exchange the floor in conversations), one of the phenomena I analyzed most in he last 15 years. Despite its relative simplicity, turn-taking confirms to be a reliable source of information.
On June 16th and 17th, the SONICOM consortium held its first face-to-face project meeting (at the National Kapodistrian University of Athens). The event involved two days of intensive discussions about audio immersive experiences from both sensory and psychological points of view. Overall, the main topic of the meeting was the estimate of Head Related Transfer Functions, i.e., that particular function that maps a sound hitting our ears into the percept we actually hear. Here are the most important lessons I learned:
The HRTF is as individual as fingerprints, it depends on the anatomy of our ears, head and torso and there are no two persons with the same HRTF;
There are two main ways to obtain the HRTF of an individual, one is to model the way perception works (through psychophysiological means), while the other is to compute the perception based on measurable physical characteristics (through data-driven approaches including machine learning);
The computation of HRTF greatly improves audio immersive experiences (an approach called “personalisation”);
There is an established community working on HRTF computation, but most of the work is based on laboratory experiments and it is unclear if and how can generalise to the “real-world”.
My role in such a setting is to see whether the perception of the artificial distance between speakers and listeners affects the perception of speakers’ personality. This would be one step in the direction of connecting the studies on HRTF to possible real-world outcomes. We are still at the beginning of the project, there are another few exciting years ahead to learn and discuss about.
I participated in an event organised by Prof Helen Minnis, a leading expert in child attachment issues, to show how Social AI can help to identify children experiencing attachment problems. The interesting connection between psychiatry and AI is that many tests used to detect mental health issues are designed to elicit behavioural variance associated to condition variance. In other words, the tests are designed in such a way that people with different underlying conditions behave in different observable ways. This is very suitable with AI approaches that, in very general terms, associate variance in observable data with variance in non-observable or non-available information.
The most interesting aspect of the event is that the audience included mostly child psychiatrists that were never exposed before to the type of AI we use to analyse behaviour. Many of them manifested surprise and mentioned that they could not imagine AI could be about this. This gives me hope that maybe one day we will bridge the existing gap between laboratory research and clinical practice.
Early 2022, Substrata gave me the chance to talk about nonverbal communication and Social Signal Processing from both scientific and technological points of view. It was a very pleasant interaction with Daniel Buchuk about all aspects of my main scientific interest, the automatic analysis of behaviour during social interactions, and a good opportunity to reflect about the business potential of my work.
I had the pleasure and honour to give a talk at the Furhat Symposium on Social Robotics:
It was an excellent mix of talks that spanned from the presentation of new social robots designed to be as easy as possible to use (Misty Robotics), to the illustration of how social robots can be used to improve education opportunities for special needs students (great talk by Axel Reitzig), to a few scientific talks given by some of the most important experts of the domain (Catherine Pelachaud, Bilge Mutlu and Kathrin Strachan). The organization was quite impressive, a further confirmation of the talent available in Furhat!
Speaking in a pub is definitely unusual, but the atmosphere was full of energy and genuine curiosity. My talk was about social intelligence in machines, a topic that always attracts attention and questions, especially about the impact that these technologies can have on the life of people. Here are the main questions I got:
Can machines be better than humans?
Is it ethical to have machines making life-changing decisions about people?
What is the role of data in Artificial Intelligence?
What is the relationship between natural and artificial intelligence?
What are the applications of socially intelligent machines?
Overall, the questions seems to account for the main worries people have about AI and it is always interesting to see that the issues remain the same across all audiences I had a chance to interact with.
I gave a keynote at the European Chatbot and Conversational AI Event:
The event involves a large number of people that look at Conversational AI from an industrial and commercial point of view. This made the Question Answering particularly interesting because all points raised by the audience focused on the very practical aspects of transferring knowledge from the laboratory to the market. Most questions revolved around the compliance with ethical requirements and current regulations in a situation that looks pretty much like a far west because the legislators tend to be late with respect to technological development.
I had a chance to present the SOCIAL CDT to the students of The Data Lab involved in a Master Program in Data Science. It was a good opportunity to learn what PhD applicants worry about:
The proposal: most students do not feel comfortable in writing a PhD proposal (not surprisingly because it is something they never did before);
The skills: there is uncertainty about the skills to be developed in order to obtain a scholarship;
Equality, Diversity and Inclusion: there are questions about possible biases in the recruitment processes.
In all cases, the best solution is to get in contact with possible supervisors. In most cases, academics are happy to answer the questions of prospective candidates and to show that applying for a PhD is not such a major challenge. The proposal is typically written in collaboration with a potential supervisor (no need to do all the work alone), the only necessary skill is openness to learn (a PhD is about acquiring new skills and not about using previously acquired skills) and all Universities are equal opportunity employers. So, if you really think PhD is what you want to do, just contact academics you like to work with and involve them in your application process!
I had the great opportunity to participate in a roundtable on data science organised by Prof Bridgette Wessels (University of Glasgow) and Prof Lesley McCara (University of Edinburgh) in the framework of SHAPE (Social Sciences, Humanities and the Art for People and the Economy). The goal of the event was to initiate a discussion about the use of data in nowadays world that included not only technological considerations (as it typically happens), but also contributions from social and human sciences. One of the main points made during the roundtable was that we need to educate data scientists to take into account ethics and Responsible Research Innovation in their work, a major challenge in an education system that tends to be highly specialistic.