Nonverbal Multimodal Social Intention Modelling

An under-explored problem in Social Signal Processing (SSP) is human social intent detection. SSP develops automated systems to interpret human social behaviour from sensors such as cameras, microphones, and wearables. Prior work estimating intent uses scenarios that are constrained enough so that the intent is determined by a congruent outcome. In reality, this is not true, otherwise emergent leaders are never female because they do not speak up and black minorities are criminals because they are often seen interacting with police. NEON addresses intention detection in more open-ended contexts involving large unstructured social gatherings such as networking or mingling events. During these events, there are no prearranged conversations, multiple conversations can occur at the same time, and all conversation come about via coordination with multiple independent actors with their own possibly conflicting goals. How do we train machines to perceive plausible intentions? Asking the individual to continuously report on their immediate intentions would contaminate their spontaneous social behaviour. NEON hypothesises that plausible social intentions are perceivable by external observers though the explanations of why (based on the behaviours of the individual and the surround social context) may vary depending on the perceivers own life experiences. To this end, NEON will develop a novel framework to harvest and learn from social intentions and their relation to future outcomes using a novel multi-task learning set up that clusters both on the traits of the perceivers as well as the observed individuals, a unified theory that situates conversations in free standing groups in both space and time and a novel Graphical Neural Network architecture to model it through multi-sensor cross modal learned neural representations. Aside from being a significant advance for SSP, NEON also benefits a range of other fields from human-robot interaction to organisational psychology