Graph schemas as abstractions for transfer learning, inference, and planning

Abstract: Transferring latent structure from one environment or problem to another is a mechanism by which humans and animals generalize with very little data. Inspired by cognitive and neurobiological insights, we propose graph schemas as a mechanism of abstraction for transfer learning. Graph schemas start with latent graph learning where perceptually aliased observations are disambiguated in the latent space using contextual information. Latent graph learning is also emerging as a new computational model of the hippocampus to explain map learning and transitive inference. Our insight is that a latent graph can be treated as a flexible template -- a schema -- that models concepts and behaviors, with slots that bind groups of latent nodes to the specific observations or groundings. By treating learned latent graphs (schemas) as prior knowledge, new environments can be quickly learned as compositions of schemas and their newly learned bindings. We evaluate graph schemas on two previously published challenging tasks: the memory & planning game and one-shot StreetLearn, which are designed to test rapid task solving in novel environments. Graph schemas can be learned in far fewer episodes than previous baselines, and can model and plan in a few steps in novel variations of these tasks. We also demonstrate learning, matching, and reusing graph schemas in more challenging 2D and 3D environments with extensive perceptual aliasing and size variations, and show how different schemas can be composed to model larger and more complex environments. To summarize, our main contribution is a unified system, inspired and grounded in cognitive science, that facilitates rapid transfer learning of new environments using schemas via map-induction and composition that handles perceptual aliasing.

• The paper introduces a graph schema approach that enables rapid transfer learning and inference by abstracting observations into latent graph structures.
• It extends the clone-structured cognitive graph model to efficiently navigate and learn new environments while addressing perceptual aliasing.
• Experimental evaluations on benchmark tasks demonstrate superior planning performance and promise for applications in robotics and adaptive systems.

Graph Schemas as Abstractions for Transfer Learning, Inference, and Planning

The paper "Graph schemas as abstractions for transfer learning, inference, and planning" introduces an innovative approach to enhancing transfer learning in AI systems by employing graph schemas. This method leverages cognitive and neurobiological insights to guide the development of AI models capable of efficient generalization with minimal data.

Conceptual Framework

The central premise is the use of graph schemas as a mechanism for abstraction. The authors propose that latent graph learning—where aliased observations are clarified in latent space using contextual data—can serve as a computational model paralleling the function of the hippocampus in map learning and transitive inference. This model treats a latent graph as a template or schema that encapsulates concepts and behaviors. By binding groups of latent nodes to specific observations, these schemas function as prior knowledge, enabling rapid learning and adaptation in novel environments.

Methodology

The paper extends the clone-structured cognitive graph (CSCG) model, which constructs graph schemas adaptable to various settings. The agent perceives its environment as a directed graph and, using this schema, navigates—and rapidly learns new environments through schema composition and map induction. The model's architecture efficiently addresses challenges like perceptual aliasing by employing a smooth probabilistic approach.

Experimental Evaluation

The efficacy of graph schemas was assessed using benchmarks such as the Memory Content Planning Game and One-Shot StreetLearn. These tasks are designed for testing rapid task-solving in uncharted environments. The results indicate that graph schemas achieve learning and planning in significantly fewer episodes than baseline methods, demonstrating strong numerical superiority. Notably, the CSCG schemas match optimal planning levels post-adaptation.

Implications and Future Directions

This research contributes a unified system built on cognitive principles and demonstrates rapid transfer learning capabilities using graph-based schemas. The implications are profound for AI applications requiring fast adaptation to new scenarios, such as robotics and autonomous systems.

Future developments could explore dynamic updating of schemas, integration with active exploration policies, and adaptation across continuous experience streams. Further research may also focus on refining schema flexibility and the maintenance of learned models.

By drawing from cognitive neuroscience and translating these insights into computational models, the work advances our understanding of efficient abstraction for transfer learning and represents a significant methodological advancement in AI research.