Published Paper Details:

: A common practice to gain invariant features in object recognition models is to aggregate multiple low-level features over a small neighborhood. However, the differences between those models make a comparison of the properties of different aggregation functions hard. Our aim is to predict the Handwritten Digit by training MNIST handwritten digit dataset by Convolutional Neural Network and Comparing the Accuracy, Loss, Validation Accuracy, Validation Loss, Time is taken, Test data Accuracy with Confusion Matrix of MaxPooling2D, GlobalAveragePooling2D, AveragePooling2D layers in our CNN model. Empirical results show that a maximum pooling operation significantly outperforms subsampling operations. We achieve 0.5-0.7% error in the maximum pooling layer and 1-0.7 % error in the Average Pooling layer while 8-9% error in Global Average Pooling Layer. While entire models have been extensively compared, there has been no research evaluating the choice of the aggregation function so far. The aim of our work is therefore to empirically determine which of the established aggregation functions is more suitable for vision tasks. Additionally, we investigate if ideas from signal processing, such as overlapping receptive fields and window functions can improve recognition performance.