ND15 / Fractional-Max-Pooling

Fractional-Max-Pooling

This repository is a crude implementation of Fractional Max Pooling written by Benjamin Graham in 2014.

Link to the paper : https://arxiv.org/abs/1412.6071 This Model is implemented on the basis of the Research paper on Fractional Max Pooling,
where the author trains the model without any training set augmentation.

Main Aspects

The main motivation of fractional max pooling is to reduce the spatial size of the image by a factor of alpha where 1 < alpha < 2. The pooling regions can either be disjoint or overlapping. The pooling regions can be defined in two ways.

• Disjoint : P = [a_i−1, a_i − 1] × [b_j−1, b_j − 1]
• Overlapping : P = [a_i−1, a_i] × [b_j−1, b_j1]
  The pooling regions can be generated either randomly or in a pseudorandom order. For generating the regions in a pseudorandom order, the sequence have to take the form of: ai = ceiling(α(i + u)), α ∈ (1, 2), with some u ∈ (0, 1).

This Repository

• Dataset is MNIST

• Here the input layer size is 28 * 28

• Architecture : 6 layers of (32nC2 - FMP(1.25)) - C2- C1 - Output, n = 1,...6

• The model is trained for 1 epoch on the MNIST dataset, where it achieved an accuracy of
  0.9540 and a loss of 0.1846.