Intrinsic-Extrinsic Coupling in Learning Dynamics
Researchers show replay can erase a classifier-head intervention's gains, formalizing intrinsic-extrinsic coupling in learning.
The paper formalizes intrinsic-extrinsic coupling, arguing that current observations do not fully determine a learner's response to further training. An executable classifier-head write preserves current logits while repairing specified historical margins. In a CLINC-derived class-incremental setting, replay changed that write's 32-update contribution from five correct predictions to zero. Nonzero interactions also appeared under output distillation, a RoBERTa backbone, and SGDW, where correct-count interactions were negative at 128 updates.
- Coupling is measured by continuation-conditioned value of a state intervention.
- Replay cut a 32-update write from five correct predictions to zero on CLINC.
- Under SGDW, correct-count interactions were negative at 128 updates.
- Guided allocation beat standard replay on cross-entropy in all five pairs.
Full article222 words · extracted from arxiv.org · click to collapse
A learner's current observations need not determine its response to further training. We formulate intrinsic-extrinsic coupling through the continuation-conditioned value of a constrained learning-state intervention, with observation-relative fibers describing present agreement. An executable finite-frame classifier-head write protects current logits while repairing specified historical margins under finite-precision acceptance checks. We distinguish local admissibility, continuation-conditioned intervention value, and complete-policy performance. A matched four-cell contrast identifies readout-specific non-additivity between the same intrinsic intervention and alternative external continuations. In a CLINC-derived class-incremental setting, replay changes the write's 32-update contribution from five correct predictions to zero. Nonzero interactions also occur under output distillation, with a RoBERTa backbone, and under optimizer-native SGDW dynamics. Under SGDW, correct-count interactions are negative in all three activated roots at 128 updates, showing that coupling need not imply positive synergy. The mathematical analysis distinguishes feasible local repairs and favorable terminal outputs from training-reachable repair regions. Separate coordination tests show that content controls match or exceed the development gain, while a five-root fresh-test comparison with Fiber present in every arm shows root-dependent rather than uniformly beneficial correct-count effects. On the secondary cross-entropy readout, guided allocation yields lower mean loss than standard replay in all five pairs. Together, these results make intrinsic-extrinsic coupling operational by connecting executable state geometry to continuation-conditioned value, matched interaction identification, and closed-loop coordination, while separating identified coupling from complete-policy performance.
Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.30185