CAI Research accepted for publication at the top journals JAIR and Neural Computation

In collaboration with the Grewe lab from University of Zurich and ETH Zurich, the Centre for Artificial Intelligence (CAI) recently published two new research articles that cover different aspects of artificial intelligence, ranging from theoretical models of image processing to the analysis of applied autonomous systems.

In the paper “The Cooperative Network Architecture: Learning Structured Networks as Representation of Sensory Patterns", published in Neural Computation, Pascal Sager and colleagues present a brain-inspired approach to visual representation learning. The proposed Cooperative Network Architecture models sensory input not as static pixel arrays but as structured networks composed of overlapping neuronal fragments. Learning is driven by Hebbian plasticity, enabling the system to capture statistical regularities in visual data. Experimental results show increased robustness to noise and an improved ability to generalize to previously unseen patterns without the use of supervised training.

In a second publication, “A Comprehensive Survey of Agents for Computer Use: Foundations, Challenges, and Future Directions", published in the Journal of Artificial Intelligence Research, Pascal Sager, Benjamin Meyer and colleagues analyze the rapidly developing field of Agents for Computer Use. These systems are designed to perform complex tasks on digital devices by interacting with software through human-like actions such as mouse movements, clicks, and gestures.

The survey provides a structured overview of existing approaches and identifies major trends in the field, including the transition from narrowly specialized text based agents toward systems built on large foundation models with visual perception capabilities. By reviewing a broad range of research papers, datasets, and benchmarks, the authors highlight key limitations of current methods, such as limited planning capabilities, fragile execution, and the lack of standardized evaluation procedures. The paper concludes with a discussion of open challenges and directions for future research.

Together, the two publications reflect the breadth of current research within the CAI, addressing both foundational questions in neural representation learning and practical challenges in the design and evaluation of autonomous software agents.