Explainable AI
This was my master thesis project. Funded by Aalto University and FCAI.
Artificial Intelligence methods, especially the fields of deep-learning and other neural network
based architectures have seen an increasing amount of development and
deployment over the last decade. These architectures are especially suited to learning
from large volumes of labelled data, and even though we know how they are
constructed, they turn out to be equivalent to black boxes when it comes to understanding
the basis upon which they produce predictions, especially as size of the
network increases.
Explainable AI (xAI) methods aim to disclose the key features
and values that influence the prediction of black-box classifiers in a manner that
is understandable to humans. In this project, the first steps are taken towards
developing an interactive xAI system that places a human in the loop; here, a user’s
ratings on the sensibility of explanations of individual classifications are used to
iteratively find Hyperparameters of the neural net classifier (VGG-16), image segmentator
(Felzenszwalb), and xAI (SHAP), to improve the sensibility of the explanations
produced without affecting classification accuracy of the classifier in the training
set. The users are asked to rate the sensibility of explanation from 1-10. The rating
from the users is fed back to the Bayesian optimization algorithm that suggests new
Hyperparameters values for the classifier, segmentator, and SHAP modules.
The
results of the user study suggests that the Hyperparameters which produced higher
ratings on explanations tended to also improve the explainability of the images,
thus generally improving the explainability for the image class. Improvement in the
out-of-sample accuracy of the classifier (for the same class) was observed in some
scenarios, but this still needs more comprehensive evaluation. More sensitive queries
for the users, explore a variety of xAI methods, a variety of datasets, as well as
conduct larger-scale experiments with users would be required to jointly improve
explanations of multiple classes.
You can find the document here.