Explaining grokking through circuit efficiency
Primary Area: visualization or interpretation of learned representations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: grokking, interpretability, generalisation, regularisation, weight decay
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Grokking occurs because the network switches from a "memorising" to a "generalising" solution, because the generalising solution is more efficient under weight decay, and we confirm novel empirical phenomena predicted by this explanation.
Abstract: We present a theory of grokking in neural networks which explains grokking in terms of the relative efficiency of competing emergent sub-networks (circuits). Grokking is an important generalisation phenomenon where continuing to train a network which already achieves nearly perfect training loss can still dramatically improve the test loss. Our theory explains why generalising circuits gradually out-compete memorising circuits. This is because memorising circuits are inefficient for compressing large datasets---the per-example cost is high---while generalising circuits have a larger fixed cost but better per-example efficiency. Strikingly, our theory is precise enough to produce novel predictions of previously unobserved phenomena: ungrokking and semi-grokking.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5816
Loading