Sequence-pooled normalization supplies global context, bypassing receptive fields

In a new arXiv paper, researcher Qing Tian overturns a core assumption in convolutional sequence labeling: that a model's receptive field defines its usable context. Tian shows that normalization layers computing statistics from the entire input sequence at inference open a sequence-spanning path that bypasses the receptive field entirely. Using a Jacobian derivation (no experiments needed for the core insight), the paper proves this path has a closed form, meaning the global summary carries precise, quantifiable information.

On a synthetic labeling task with computable optima, a network reaching only 9 positions comes within 0.009 of the whole-sequence optimum, while a standard receptive-field-only baseline performs near chance. The effect holds on real data too: closing the path by computing per-position statistics multiplies the value of enlarging the receptive field by up to an order of magnitude on simulated genomes and 1000 Genomes haplotypes. The finding also exposes a major flaw in common ablation studies—severing receptive-field-enlarging blocks overstates their contribution 8.3–16.1-fold compared to retraining from scratch. The effect diminishes where labels switch frequently, but for long runs of repeated labels, normalization is doing far more heavy lifting than previously credited.