Emotion Recognition: A Statistical Learning Approach

Statistical learning refers to the ability to learn through the discovery of patterns and structures.

I propose to investigate emotion recognition using a statistical learning perspective in order to understand (i) why some emotions are harder to recognise than others; and (ii) why individuals with autism spectrum disorder (ASD individuals) have more difficulty recognising emotions than neurotypicals (i.e., individuals without autism).I argue that part of the difficulty in recognising certain emotions lies in how reliable or consistent the auditory and visual cues are in signalling the emotion.

That is, if particular cues consistently signal or have a high probability of signalling an emotion (e.g., 'happy' is consistently signaled by squinty eyes and grin/smile), then that emotion would be easier to recognise than emotions that are signalled by inconsistent cues (e.g., sarcasm may have varied expressions depending on the individual, context, etc. and so sarcasm would be more difficult to recognise).

To investigate this, I will use an audio-visual emotion database that is currently under development to quantify the variability of cues across speakers in signalling the intended emotion.I propose that the difficulty ASD individuals have with recognising emotions lies in a general difficulty with consolidating probabilistic information.

In terms of emotion recognition, this would manifest as a difficulty with making a correct inference of the intended emotion given particular cues, which vary in their probabilities in signalling the emotion.

To investigate this hypothesis, I will conduct a behavioural and a neural experiment comparing ASD individuals with neurotypicals on probabilistic learning to determine whether group differences exist and whether probabilistic learning is related to emotion recognition.Outcomes of this project may inform intervention practices for ASD individuals and provide a general framework of understanding other ASD characteristics.