Reproduce “Neural Network based Formation of Cognitive Maps of Semantic Spaces and the Emergence of Abstract Concepts” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Paul Stoewer, Achim Schilling, Andreas Maier, Patrick Krauss
Researchers at the University Erlangen-Nuremberg developed a neural network model utilizing multi-scale successor representations to construct cognitive maps of abstract semantic spaces, demonstrating how the discount factor regulates map granularity and facilitates the emergence of conceptual hierarchies. This approach also enables robust inference of missing features from partial input, mimicking the brain's capacity to derive context from incomplete information.
Key insights
- The discount factor (γ) within the SR framework directly dictates the scale of the cognitive map, providing a computational explanation for the hierarchical organization of knowledge from fine-grained details to abstract concepts.
- Successor representations can function as 'weighted pointers' to past memories, integrating aspects of the multiple trace theory of memory with cognitive map theory.
- The model demonstrates that context and missing features can be robustly inferred from novel or incomplete semantic inputs by referencing the learned cognitive map.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.