Standard deep vision networks have run on addition since ResNets made depth practical in 2015 — a new module argues multiplication works better.
Researchers have proposed PURe, a Product-Unit Residual module that replaces the additive local transformations common to modern vision architectures with multiplicative ones. The core problem with product units in deep networks has always been optimization instability; PURe sidesteps this with a log-domain formulation that keeps multiplicative aggregation numerically well-behaved inside residual hierarchies. The module is designed as a drop-in replacement for standard residual units — no architectural rebuild required. Evaluations on ImageNet, CIFAR-10, and Galaxy10 DECaLS showed moderately deep models with PURe matching or beating substantially deeper ResNet baselines at smaller parameter counts; on the AMOS medical imaging benchmark, the module also improved CT slice segmentation under 3D case-level evaluation.
The accuracy-to-parameter trade-off claim is the one worth scrutinizing. Smaller models that perform comparably to deeper ones have obvious appeal for medical imaging and edge deployment, where compute and memory carry real cost — and the AMOS result is a harder benchmark to hand-wave than ImageNet alone. The "drop-in" framing also lowers the adoption bar considerably compared to architectural overhauls.
That said, the history of plug-and-play residual improvements is full of benchmark gains that don't survive contact with production workloads or datasets outside the test suite. The log-domain trick is clever, but whether multiplicative interactions generalize as cleanly as addition across the full range of vision tasks remains an open question.