Lecture 11: Informed Decision-Making

Human-Centred AI grand challenges

machine teaching, human in the loop, human sensing, human supervision, security, privacy

Clinical decision support challenges

limited data, fragmentation, personalisation, ethics and safety, sensitive personal data

Human in the loop

we humans interacting with AI and by this way we can overcome the fragmented nature of the system

informed clinical decision-making

you need to be able to provide a reason for the probability that was determined

Explainable predictions

Shapley Additive exPlanation

based on cooperative game theory

shapely values assign a value to each feature based on the feature’s contribution to the outcome of the model

Shapley exPlanation Importance

we must explain the functionality of the Ai to the user in the user interface

Lime Technique

having one dot in the space, and then creating many points around that point, this one dot is the input and the many points surrounding it are the many output probabilities; the coefficient of this linear model are the feature importances

Lime Model Issue

there is a random procedure in the pipeline, you are suffering from robustness; every time you are performing the same experiment, you will not have the same results

Crowdsourcing

nota

Counterfactual explanations

used to generate feature values, we are trying to identify which is the minor change that we can do in the input space in order to change the prediction from +ve to -ve (or vice versa); since I will be able to identify this minor change that has a very large effect on the final outcome it means that this feature plays an important role

Example of Human in the loop

data labelling in audio recordings of coughs to understand the symptoms