How can you apply explainable AI research to your career?

To apply Explainable AI (XAI) in your career, follow these steps: 1. Build XAI Methods: If you're an AI developer, create algorithms that clarify how AI decisions are made. Think of ways to show what's happening inside the model visually. 2. Choose XAI Tools: As a user, opt for AI systems with built-in explainability. Look for platforms that offer understandable insights into how decisions are made. 3. Assess Impact: If you're a stakeholder, use audits to check if the AI is fair, safe, and transparent. Include diverse opinions in the AI's design and rollout. By integrating XAI, you'll help make AI more understandable and trusted.

I am curious how this is fanning out in the security world to check against security measures etc. I was former military working with IT security but not involved with the AI / ML evolution regarding security audits, challenges etc.

When it comes to our relationship with technology, the concept of “Trust” is often the first thing that comes to mind. To trust the results of an AI model, whether it’s making predictions or classifications, we need to ensure that the model employs certain techniques. This is where XAI comes into play, fostering trust by explaining the model’s decisions. Whether it’s making decisions based on predictions that impact human safety, or understanding and detecting bias in these decisions which could lead to unfairness between discriminative characteristics such as gender or race, XAI provides the transparency we need to trust AI model.

Some XAI methods combine post-hoc and intrinsic explanations. For instance, hybrid models may use symbolic reasoning to provide high-level explanations while also generating post-hoc explanations based on feature importance to offer more granular insights. The aim of these approaches is to leverage the strengths of both methods.

Another method to add that can be utilized for XAI is SHAP (SHapley Additive exPlanations). This technique is a model-agnostic post-hoc explainability method that employs coalitional game theory, specifically Shapley Values, to estimate the significance of each feature.

As Goebel et al. 2018 pointed out in their paper “Explainable AI: The New 42?”, the complexity of AI often makes it difficult, if not impossible, to explain. Thus, selecting appropriate methods for XAI depends on other approaches. Consider, for instance, application of XAI in image-based visualization classification. An approach known as “deep visual explanation” could be utilized. This method is useful in application where learned models for image classification need to be debugged. It involves creating a heat map filter to pinpoint where in the image a classification decision was made. A method such as Kullback-Leibler (KL) divergence gradient can then be used to identify differences in learned variable distributions to create these maps.

In research paper titled “Towards a Rigorous Evaluation of XAI Methods on Time Series”, Udo Schlegel and his team presented two sequence verification methods and a methodology for automatically evaluating and testing XAI explanations on time series. They propose two novel techniques for testing and evaluating XAI methods for Sequence Evaluation: Swap Time Points and Mean Time Points. On the other hand, the proposed methodology is conducted out in three stages: 1) model training and evaluation, 2) creation of model explanations, and 3) evaluation and verification of explanations.

Here are some interesting research papers that delve deeper into the method of Explainable AI (XAI): - “Explainable AI: A Review of Machine Learning Interpretability Methods” by Linardatos P, Papastefanopoulos V, Kotsiantis S, published in the Journal of Entropy and Information Studies. - "A Review of Trustworthy and Explainable Artificial Intelligence (XAI)” by V. Chamola, V. Hassija, A. R. Sulthana, D. Ghosh, D. Dhingra and B. Sikdar, featured in IEEE Access. - “Unraveling ML Models of Emotion With NOVA: Multi-Level Explainable AI for Non-Experts,” by A. Heimerl, K. Weitz, T. Baur and E. André, published in IEEE Transactions on Affective Computing, vol. 13.