abstratos

Abstratos is a theoretical framework in cognitive science and artificial intelligence that describes a method for constructing higher-level abstract representations from lower-level data. The core idea is to separate concrete instances from the abstract concepts that generalize across instances, enabling models to transfer knowledge across tasks and domains with limited data.

Origin and terminology: The term abstratos was coined in 2023 by researchers pursuing a formalization of abstraction processes. The name combines a root from Latin abstrahere, meaning to draw away, with a suffix to denote a process or state.

Concept and structure: A typical abstratos model comprises four elements: an abstraction layer that defines a space of abstract concepts; a mapping function that converts raw inputs into this space; invariants that preserve essential structure under transformations; and an evaluation criterion that balances predictive performance with interpretability and robustness.

Operationalization: In practice abstratos is implemented as a modular pipeline in which an encoder projects data into an abstract space, a symbolic or rule-based module manipulates these abstractions, and a decoder or planner uses the results for downstream tasks such as classification or decision-making.

Applications and reception: Proponents argue that abstratos improves sample efficiency, enables zero-shot generalization, and enhances explainability by revealing task-relevant concepts. Critics caution that the choice of abstractions can be subjective, risk information loss, and add computational complexity. Ongoing work explores formal guarantees and benchmarks for abstraction quality.

See also: Abstraction, Representation learning, Symbolic AI, Ontologies.