Visual Network Representation Extension for a New Artificial Learning Methodology on Behavioral Agents

Contemporary machine learning methods such as neural networks continue to face persistent challenges, including catastrophic forgetting (destruction of past knowledge when learning something new) and the lack of decomposable, engineerable, and well-structured internal representations. A recently proposed artificial learning methodology [1], termed varsel networks, offers a promising path to overcoming these issues.
 
So far, this methodology has been demonstrated in two separate domains: (1) behavioral agents operating on simple, discrete state spaces (lists of state variables), and (2) a proof-of-principle variant for simple visual tasks such as image classification, where images are represented as graphs (networks) of features and their positional interrelations.
 
To extend varsel networks to critical domains such as behavioral agents acting on visual inputs (e.g., video games, robotics), these two strands of work (the behavioral variant and the visual-graph variant) must be integrated. The aim of this project is to realize an initial prototype of such an integration. Your will take a simple 2D visual environment in which an agent can act, develop a visual network representation for that environment (features and their interrelations), and apply varsel networks to learn a model of the environment based on this representation.
 
Prerequisites: Fluency in Python. Prior familiarity with computer vision is a plus, but not required.
 
[1] Erden, Z. D. (2025). Foundations of a New Learning Paradigm in AI Grounded in the Principles of Evolutionary Developmental Biology, Part III: Varsel Networks (Doctoral dissertation, EPFL).