Advancing Pooling Techniques

The possibility of fitting distributions uncovered a myriad of pooling statistics. It is also made possible using advanced techniques, such as an adaptive selection of distribution. Such techniques have significance because optimal pooling depends on the characteristics of feature maps in convolutional networks and the dataset.

Automatically determining the optimal pooling is challenging. Moreover, MLEs are learned to be the appropriate pooling statistic. But they’re unavailable for some distributions.

Besides, convolutional feature maps are likely to depend on most datasets because nearby features are correlated. However, pooling statistics are developed assuming their independence. This section provides a few directions to address these challenges.

Adaptive distribution selection

Feature maps for different samples in a dataset can differ significantly. For explanation, the illustration below shows images with an object at the center and corner, respectively. The images are filtered through a Sobel filter. Shown in the illustration is the feature map distribution yielded by the filter.