Design by Growth: Reading Out Natural Attention from a Memory Grown from Basic Elements
Memory components for neural networks are usually designed first and trained afterwards. We propose design by growth. We build a task world that requires a capability, give a recurrent network only basic elements, and let training find a solution. We then read the solution out of the trained parameters and write it as a component. We apply this to working memory. The task is to store several pairs of a cue and a content word and, after a delay, to output the content word for a given cue. A network made only of recurrent cells holds just one item. We add fast weights between the cells, which are connections that change while the network runs, and let each cell learn how much it takes part in storing and retrieving. With these elements, the network grows a memory that holds several items. Its parameters show three rules. Storing and retrieving use the same key, carried by a few cells that track the most recent cue. The content is held by the other cells. Writing happens only when content arrives. We write these rules as a component that we call natural attention. In one version, the fraction of old content erased before each write is learned at every step; in the other, all old content is erased. In the task world, both versions need about one eighth of the training that the grown network needs. On multi-query associative recall (MQAR), they need half the training steps of Gated DeltaNet and about one fiftieth of those of attention. On selective copying, they exceed 99% accuracy, while Gated DeltaNet reaches about 30%. These rules were not designed in advance; they were read out of a memory that the network grew.
Authors
- Wenyuan Li (ORCID: https://orcid.org/0009-0004-2391-8508)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-14
- DOI
- https://doi.org/10.5281/zenodo.24224519
- Primary Topic
- Neural Networks and Applications
- Type
- preprint