F21ZK - Dependable and Deployable Artificial Intelligence for Robotics 1

Aims

This course aims to:

Introduce students to the principles of specification and implementation of autonomous systems that have machine learning/AI components.
Ensure that CDT-D2AIR educates a cohort of specialists who understand both the complexities of autonomous system specification and verification, and existing solutions and trends.
Expose the students to advanced principles of specification-based design. Systems engineering is a vital discipline in the development of dependable and deployable systems, with a core element being the creation and adherence to a clear specification.
Expose students to the state-of-the-art in modern verification technologies, including the tools for verification of machine learning and hybrid systems (e.g., SMT solvers, model checkers, functional programming languages, neural network verifiers).
Ensure that students understand the complexities of verification of autonomous systems with machine learning components. In particular, that they know and understand algorithms and tools that provide the supporting infrastructure for verification, i.e., specification-driven training (as opposed to data-driven training), automated ways to connect verification of machine-learning components with verification of the overall system. Students will learn how to apply these methods to various aspects of robotics, throughout the design and development pipeline.

Syllabus

1. Study principles of specification-based design (1.1 The students will study methods for gathering requirements and specifications for design of autonomous systems)

2. Study principles of autonomous systems design (2.1 The students will learn to understand major pitfalls and solutions for designing complex autonomous systems with machine learning components)

3. Practice and master Machine-Learning and AI methodologies in modern autonomous system design (3.1 Machine-learning is deployed for sensor processing, controllers, and user interaction in complex autonomous systems. All three are substantially different, on the technical and systems side. The students will learn to identify each group of methods.)

4. Understand History and state-of-the-art in software and hardware verification (provers, solvers, model checkers) (4.1 Verification domain has a long history in Computer Science, and major break-throughs. Understanding its main methods, succesess and pitfalls is the first step to applying the methods for verification of autonomous systems. The students will learn to identify the methods and appreciate their scope.)

5. Master State-of-the-art tools in Neural Network verification, such as Marabou and Alpha-Beta-Crown (5.1 Verification of neural networks can be done by employing tools like Marabou and Alpha-Beta-Crown. The students will practice to apply these tools.)

6. Learn to use machine-learning methods for property-driven training (and their role in data-driven training), such as differential logics and adversarial training (6.1 Often, machine learning and optimisation need to be deployed in order to ensure that models adhere to stated verification properties. The students will learn about existing methods that achieve this.)

7. Study challenges and emerging trends in verification of cyber-physical systems and autonomous systems with machine learning components (7.1 Verifying cyber-physical systems with machine learning components is a nascent field, with several research groups competing for best solutions and tools. The students will be taught to understand this landscape.)