- New module lets existing CNNs use multiplicative interactions.
PURe introduces a 2‑D product unit that works in the log domain, sidestepping the instability that has kept multiplicative layers out of deep nets. The authors drop it into ResNet‑style blocks and test on Galaxy10 DECaLS, ImageNet and CIFAR‑10, as well as slice‑based CT segmentation on the AMOS benchmark. Across the board the modified networks either match deeper ResNets with half the parameters or exceed them outright.
The significance lies in proving that explicit multiplicative aggregation can be a practical design primitive. It narrows the accuracy‑parameter gap, which could matter for edge devices and medical imaging pipelines that cannot afford large models.
So far the gains are modest but consistent, suggesting a new lever for architects tired of squeezing performance from additive layers alone.