Interpretable Patterns in Random Initialization Unveil Final Representation
Keywords: mechanistic interpretability, training dynamics, science of deep learning, representation learning
TL;DR: We define some "completeness" metrics for partially formed representations in initial embedding and evaluate on three toy models how well they predict final representations.
Abstract: The field of mechanistic interpretability has made strides in unraveling models' hidden representations but is often puzzled by why specific representations form. This paper addresses a crucial question on this front: when a neural network can learn multiple distinct representations to solve a task, how does it "choose" among them during training?
We suggest that, at initialization, instead of starting from an empty scratchpad, the model's embedding already contains partially formed representations of varying ''completeness.'' Models tend to develop a representation that is more "complete" at initialization, disregarding less complete alternatives.
We empirically examine this hypothesis on algorithmic toy models with clearly defined final representations from which we can elicit an interpretable signal to evaluate such "completeness" of possible representations in the initial embedding. We find that the representations with high initial signals are chosen by the model with high probability, a pattern consistent across models with a single learned representation (remainder equivalence, multi-digit XOR) and with multiple, redundant representations (modular addition).
Finally, we investigate the role of embedding dimensionality on model's representation and their ``completeness.''
Our results with toy models show that the seemingly chaotic initialization contains many interpretable patterns to understand the training dynamics of representations.
Primary Area: interpretability and explainable AI
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 12814
Loading