The full version of the thesis will be available soon. 

Our lives are increasingly becoming digitized, and data is being produced like never before. So much so that most of the data will never be seen by human eyes. Instead, artificial intelligence (AI) algorithms analyze our data, extract meaningful information, and make critical decisions that impact our everyday lives. Unfortunately, not all AI is created equal. Deep learning (DL) is a popular and powerful type of AI loosely inspired by how brains work. DL models are very complex, and it is difficult to scrutinize model decisions. However, DL is widely used in real-world applications, and we need to trust the model’s decision-making process. That is where explainable AI (XAI) comes in. The thesis starts from a high level and zooms into applications of explainability.

First, four components are considered that lead to most issues regarding XAI: Users, laws & regulations, explanations, and algorithms. I find that concerns regarding the scrutability of data use, inference quality, and adversarial attacks can be addressed with XAI, while concerns regarding flawed data collection and undesirable outcomes cannot.

Next, I take a closer look at the explanation algorithms themselves, creating an extensive taxonomy surveying more methods, and I consider issues related to the algorithms, such as bias in the data and model debugging. I conclude that a lot more has to be done in terms of XAI for the end-users. The evaluation of XAI methods themselves is critical and can benefit from more attention.

In the following two papers, I dive deeper into the mechanics of 3D convolutions (3DConvs), a popular DL model component that allows for simultaneous spatial and temporal feature extraction from video data. However, 3DConvs do not easily lend themselves to scrutinization because the temporal and spatial domains are difficult to separate. I develop a novel variation on the 3DConv that separates the temporal and spatial domain. I experiment with 3TConvs and explore their use in XAI.

Finally, I study the effects of background on personality perception by DL algorithms. In this paper, aspects discussed in previous chapters are observed in a practical setting: A significant component of this work is the incorrect use of the dataset in previous literature investigating the role of the background. First, the training data contaminated the test and validation splits, and second, previous studies included information about the face in the background when they should not have. Additionally, I demonstrate how to use perturbation-type XAI methods without introducing out-of-distribution data to the DL model.

The thesis closes with the conclusion that there is no one-size-fits-all solution. Instead, each situation requires its own approach where the entire context of explainability is taken into account. In particular, the end-user needs more attention as currently, there are almost no methods that explain algorithm decisions to the end-user. Also, the usefulness of explanations for the end-user needs to be evaluated. Finally, the explanation algorithms themselves need to be evaluated in terms of correctness and usefulness. The only thing worse than no explanation is the wrong explanation.