DISL-Lab (KI-Labor Dependable Intelligent Systems)

BMBF-Project DISL-Lab (Dependable Intelligent Systems)

Description

The primary research focus for the CogSys Team in this project is the development of novel methods for explainable reinforcement learning. While Explainable AI (XAI) has seen a sharp rise in popularity for supervised and unsupervised machine learning tasks, its occurrence in the reinforcement learning domain is still much sparser. However, given the potential impact of autonomously and intelligently acting machines in the future – the ultimate goal of reinforcement learning – we are expecting to see an increasing need for transparent methods in this area of AI as well. 

As of today, the dangers of non-transparent AI are already apparent and frequently experienced in particular by disadvantaged members of our society. This adds to the issue that safety guarantees for intelligent agents – think self-driving cars or robots – are currently very hard or even impossible to obtain, let alone prove equal safety for all societal groups. A complete understanding of an agent’s rationale will thus be inevitable in order to ensure fair and safe actions by advanced AI in the future.

Besides the ability to validate the safety, fairness and rationality of explainable agents, we are also interested in leveraging their transparency in order to improve and accelerate the learning process. Given that reinforcement learning currently requires large amounts of data or long-running simulations of artificial environments, we aim to provide relevant human knowledge to an agent who might otherwise require much longer to obtain such knowledge itself.

 

Methodology

The following list provides an overview of methods we are considering:

Leveraging existing XAI methods to reinforcement learning: While there are important differences between (un-)supervised machine learning and reinforcement learning, there exist also many parallels that connect both domains. Therefore, we see it as a logical step to take a closer look at existing methods like LIME or SHAP and examine their applicability to reinforcement learning. 

Improving inherently transparent methods: Even today, we already have access to transparent methods that have been developed over many decades of machine learning research. Commonly known frameworks like linear regression and decision tree models from the statistical learning domain or inductive logic approaches from symbolic AI all share a considerable degree of transparency. However, it is commonly agreed on that these methods in their original form struggle to deal with nowadays’ complex vision, textual or merely big data as compared to popular deep learning frameworks. Nevertheless, we see a big potential in advancing these methods to be able to solve these tasks while retaining their capability to be understandable by humans.

Differentiable Programming: Initially, this term was jokingly introduced by Yann Lecun referring to the complexity of modern deep learning algorithms. However, today’s successful machine learning models tend to look more and more like actual computer programs turned end-to-end differentiable. Automatic Differentiation frameworks like Tensorflow or PyTorch allow for mathematically expressing abstract programs and fine-tuning their parameters in a machine learning sense. This obviously leads to the option of writing highly transparent programs that a human can make sense of and improve upon.

 

Activities

  • “Dependable AI under heavy tailed distributions” - Short presentation at KI2020, Workshop Dependable AI
  • “Deep Learning – Das next big thing oder die nächste große Blase?“ – Presentation at Total Digital 2019, Coburg
  • “Wie lernen Maschinen? – Ein- und Ausblicke in aktuelle Erkenntnisse aus der KI-Forschung“ – Presentation at Total Digital 2020, Coburg

 

Publications

 

Supervised Theses