You are here

Home › Research › Projects ›

Junior Research Group Correctable Hybrid Artificial Intelligence (chAI)

A project funded by the German Federal Ministry of Education and Research, funding code 01IS24058

Contact: Dr. Gesina Schwalbe (group lead)

Motivation, Background, and Problem Statement

Data-driven artificial intelligence (AI) methods, especially deep neural networks (DNNs), hold great potential for automating processes in safety-critical applications. Examples include environmental perception, such as camera-based perception in robotics or autonomous driving. However, to ensure the safe application of these methods, their verifiability and correctability must be guaranteed. This also applies to symbolic knowledge, i.e., knowledge describable in natural or formal language, such as traffic rules ("If the light is red, stop") or known object relationships ("A head belongs to a person"). Unfortunately, current DNNs fail to meet these requirements: these purely statistical models are large and opaque, offering insufficient guarantees for verifying learned knowledge or integrating and correcting symbolic knowledge effectively.

Objective and Approach

The junior research group Correctable Hybrid Artificial Intelligence (chAI) aims to develop methods that allow symbolic knowledge to be (1) verified, (2) specifically corrected within the DNN (→ correctable), and (3) incorporated during training and design (→ hybrid). This will involve leveraging methods

  • from the field of explainable AI to uncover and track information within DNNs, in particular introspection techniques like concept activation vectors;
  • for formal solvers to obtain guarantees;
  • the field knowledge representations.

Project Structure

The project will explore three different intervention points for integrating knowledge into a DNN across three doctoral research projects:

  1. Through information in the DNN’s intermediate outputs (e.g., “Does the DNN ‘know’ what a head is?”),
  2. Through the internal structure of the DNN’s components and connections (e.g., Where and how can traffic rules be represented in the connections?), and
  3. Through the hybridization with classical symbolic algorithms or knowledge bases (e.g., How can a DNN serve as a language interface for classical robot planning?).