Geometric Mean Pooling for Equal-Weight Multiplicative Coarse-Graining
The paper proposes pooling by multiplying signs and taking the geometric mean of feature magnitudes.
The authors frame Geometric Mean Pooling as a non-learned alternative to average and max pooling for cases where multiplicative composition matters. It preserves a global multiplicative statistic under non-overlapping hierarchical pooling. In tests, it beats average and max pooling on synthetic product-signal tasks, but its results on images and molecular lipophilicity depend on representation and where pooling is applied. ArXiv · AI/CL/LG's note
The authors frame Geometric Mean Pooling as a non-learned alternative to average and max pooling for cases where multiplicative composition matters. It preserves a global multiplicative statistic under non-overlapping hierarchical pooling. In tests, it beats average and max pooling on synthetic product-signal tasks, but its results on images and molecular lipophilicity depend on representation and where pooling is applied. ArXiv · AI/CL/LG's note
score 4