Experience as a Persistent Object: Online Temporal Credit in DABSN
This paper introduces and rigorously characterizes a persistent experience object in DABSN that preserves temporal learning information beyond computation graphs. Mathematical proofs and extensive experiments establish its distinction from full parameter sensitivity, its causal editability, and its ability to support objective-conditioned continuation without external targets. A unified learning mechanism uses the same object for credit assignment, update direction, and learning pace without persistent optimizer state in the core, achieving performance competitive with AdamW and backpropagation through time on demanding sequence tasks.
Authors
- Nicholas Rosdahl
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24268686
- Primary Topic
- Neural Networks and Applications
- Type
- preprint