Experimental neural learning research for FlossWare.
This repository explores small, understandable learning systems before introducing framework or infrastructure complexity.
The initial question is deliberately small:
Can a simple learner acquire a bounded behavior from examples, retain what it learned, and generalize to examples it did not see?
The experiments will examine:
- learning versus memorization
- retained learner state
- generalization
- deterministic evaluation
- persistence and restoration
- increasingly complex learning rules
- interaction among multiple learners
Core learning mechanisms are intentionally implemented from first principles where practical. External libraries should earn their place by enabling an experiment, not by defining the experiment.
- neural: learning mechanisms and learner behavior.
- neural-ai: experiments where AI models act as teachers or knowledge sources.
- loom: language-neutral execution substrate, introduced when experiments need it.
- loom-ai: AI execution contracts and model-provider semantics.
neural does not depend on neural-ai. A future integration should pass teaching signals into a learner through a small, language-neutral boundary.
Experiment 001 defines the first research target: a minimal binary learner evaluated on held-out examples.
No Maven project, framework, or Loom dependency is required at this stage. The point is to learn what the learner needs before building a small software cathedral around it.
Each experiment should state:
- hypothesis
- smallest implementation needed
- training data
- evaluation data
- measurements
- observed result
- what the result changes about the next experiment
Results should be reproducible where practical, and negative results are useful evidence rather than failures to hide.