Mechanistic Mode Connectivity
Ekdeep Singh Lubana,u00a0Eric J Bigelow,u00a0Robert P. Dick,u00a0David Krueger,u00a0Hidenori Tanaka
We study neural network loss landscapes through the lens of mode connectivity, the observation that minimizers of neural networks retrieved via training on a dataset are connected via simple paths of low loss. Specifically, we ask the following question: are minimizers that rely on different mechanisms for making their predictions connected via simple paths of low loss? We provide a definition of mechanistic similarity as shared invariances to input transformations and demonstrate that lack of linear connectivity between two models implies they use dissimilar mechanisms for making their predictions. Relevant to practice, this result helps us demonstrate that naive fine-tuning on a downstream dataset can fail to alter a modelu2019s mechanisms, e.g., fine-tuning can fail to eliminate a modelu2019s reliance on spurious attributes. Our analysis also motivates a method for targeted alteration of a modelu2019s mechanisms, named connectivity-based fine-tuning (CBFT), which we analyze using several synthetic datasets for the task of reducing a modelu2019s reliance on spurious attributes.
