FOR 5368:  KI-FOR Abstract Representations in Neural Architectures (ARENA)

• Subject Area: Computer Science, Systems and Electrical Engineering ; Medicine ; Social and Behavioural Sciences
• Term: since 2022
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 459426179

The overarching goal of the research unit ARENA is to advance understanding of how semantic knowledge representations can emerge and are encoded in neural architectures – in human brain networks and in artificial intelligence (AI) models. We will study knowledge representations across varying levels of abstraction, ranging from invariance against different sensory instances of the same object (e.g., seeing an object from different angles) to categorical representations summarizing commonalities across different objects, to linguistic meanings denoting such objects and abstract conceptual-semantic representations that can be accessed through multiple input modalities (like the concept ‘baby’ being activated by seeing the image as well as by hearing a baby crying or reading the word ‘baby’). Understanding the coding and computations underlying such representations remains a challenge for cognitive neuroscience and AI modeling alike. We propose that AI models in combination with machine learning and cognitive computational neuroscience expertise can be harnessed to systematically extend our understanding of hierarchically organized knowledge along such gradients of abstractness in the human mind and brain, and at the same time serve as inspiration for AI model development. To this end, ARENA will integrate the long tradition of experimental psychology and cognitive neuroscience with the explosive development in AI-oriented computer models in recent years to tackle the following research questions: (1) What are the computational principles underlying the representation of increasingly abstract semantic knowledge in the brain and how can these be integrated into AI models? (2) What are the principles that allow for abstract representations to emerge in the developing brain and how can these be incorporated into AI models? (3) How can the brain learn such representations so much more autonomously than current AI models and how can we mimic this? (4) How can these high-dimensional representations of abstract knowledge in the brain and in AI models be accessed and navigated efficiently for behavior? Deriving computationally explicit accounts of semantic representations will bring us closer to deciphering cognitive processes in the human brain and may pave the way for more versatile and human-like AIs. To achieve this ambitious goal, ARENA unites PIs from the fields of computer science, psychology, and cognitive neuroscience, including a bridge professor at the intersection of these fields. Combined, ARENA PIs have extensive experience in AI and other computational modeling approaches and long-standing experience in experimental investigations of human behavior and brain processes. The interdisciplinary nature of the ARENA research unit will serve as catalyst for aligning common research interests across disciplines, and to establish formal structures that allow us to systematically cultivate interdisciplinary synergies.

• DFG Programme: Research Units
• International Connection: USA

Projects

• Autonomous Learning of Object Representations (Applicant Triesch, Jochen )
• Bridging Levels of Abstraction in Brains and Natural Language Processing Machines (Applicant Toneva, Ph.D., Mariya )
• Coordination Funds (Applicant Fiebach, Christian )
• Emergence of Abstract Representations in Contextualized Multimodal Models (Applicant Roig, Gemma )
• Hierarchical Organization of Distributed Semantic Knowledge in the Human Language System (Applicant Fiebach, Christian )
• Navigating Knowledge Spaces Using Cognitive Maps (Applicant Kaschube, Matthias )
• The Emergence of Abstract Representations in Learning and Development (Applicant Shing, Yee Lee )
• The Hierarchical Structure of Abstract Scene Representations (Applicant Vo, Melissa Le-Hoa )

• Spokesperson: Professor Dr. Christian Fiebach