Ryan Davies and Sam Rupprechter
Tue 07 Nov 2017, 11:00 - 12:00
IF 4.31/4.33

If you have a question about this talk, please contact: Gareth Beedham (gbeedham)

Ryan Davies

Title: Providing Feedback to Deep Networks via Interpretable Representations

Abstract: Current deep networks can be complex, making it difficult to understand why the model makes the decisions it makes. There is recent work on finding ways to explain the output of these complex models in an interpretable way, that can be understood by people. Interpretable representations can help people to understand why deep networks make the predictions, but provide no direct guidance of how people should fix the model. We propose a mechanism for feedback to allow this, in which a person can provide a desired interpretable representation to repair incorrect predictions. The model is retrained to match both the original training data and the feedback. I will talk about the importance of interpretability, and why it would be useful to be able to provide feedback to complex models using interpretable explanations. I will also talk about some of our current work which aims to implement feedback.


Sam Rupprechter

Title: Value-based decision-making impairments in Major Depressive Disorder

Abstract: In recent years the field of computational psychiatry has emerged, which aims to use mathematical and computational techniques to better understand mental illnesses such as depression.  Although Major Depressive Disorder (MDD) is one of the most common psychiatric disorders in almost every country, and it's prevalence is increasing especially in highly developed countries, the disorder is still poorly understood.  Depressed patients suffer from debilitating symptoms and also display impairments in various (value-based) decision-making tasks, for example using Pavlovian or instrumental conditioning paradigms.

Behaviour in these tasks is often well described using simple Bayesian or reinforcement learning models.  I will present results from a Pavlovian conditioning experiment probing differences in decision making between depressed patients and healthy controls.  Patients showed overall worse decision making and computational modelling revealed that across groups the inability to follow internal value estimations correlated with neuroticism, which is a personality trait strongly linked to a multitude of mental disorders and a promising endophenotype for depression.