Interpretation of intelligence in CNN-pooling processes: a methodological survey

The convolutional neural network architecture has different components like convolution and pooling. The pooling is crucial component placed after the convolution layer. It plays a vital role in visual recognition, detection and segmentation course to overcome the concerns like overfitting, computation time and recognition accuracy. The elementary pooling process involves down sampling of feature map by piercing into subregions. This piercing and down sampling is defined by the pooling hyperparameters, viz. stride and filter size. This down sampling process discards the irrelevant information and picks the defined global feature. The generally used global feature selection methods are average and max pooling. These methods decline, when the main element has higher or lesser intensity than the nonsignificant element. It also suffers with locus and order of nominated global feature, hence not suitable for every situation. The pooling variants are proposed by numerous researchers to overcome concern. This article presents the state of the art on selection of global feature for pooling process mainly based on four categories such as value, probability, rank and transformed domain. The value and probability-based methods use the criteria such as the way of down sampling, size of kernel, input output feature map, location of pooling, number stages and random selection based on probability value. The rank-based methods assign the rank and weight to activation; the feature is selected based on the defined criteria. The transformed domain pooling methods transform the image to other domains such as wavelet, frequency for pooling the feature.