Active perception and disentangled representations allow continual, episodic zero and few-shot learning
David Rawlinson, Gideon Kowadlo · 2026
Read Active perception and disentangled representations allow continual, episodic zero and few-shot learning by David Rawlinson, Gideon Kowadlo (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-19355.txt.
Description
Generalization is prized in machine learning, but perhaps not every part of a system needs it. David Rawlinson and Gideon Kowadlo describe a Complementary Learning System whose fast learner gives up generalization entirely, trading it for continual learning from zero or few examples. Standard training tends to entangle representations at entity boundaries, causing destructive interference when rapid, large updates are needed.
Unlike most such systems, which use episodic memory mainly for replay, their fast, disentangled learner works as a parallel reasoning system. A contextual bias from the fast learner nudges the slow, statistical learner to encode new stimuli in familiar, generalized terms, helped by active perception. The result shows fast, context-driven reasoning coexisting with slow, structured generalization.
Catalog record
| CREATOR | David Rawlinson, Gideon Kowadlo |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-19355