[Paper Review] Fractional Max-Pooling

Benjamin Graham|arXiv (Cornell University)|Dec 18, 2014

Advanced Neural Network Applications14 references335 citations

TL;DR

This paper introduces fractional max-pooling (FMP), a stochastic spatial pooling method that reduces feature map size by non-integer factors (1 < α < 2), enabling finer-grained feature hierarchy and improved invariance. FMP reduces overfitting and achieves state-of-the-art performance on CIFAR-100 without dropout, outperforming standard 2×2 max-pooling across multiple datasets including MNIST, CIFAR-10, and handwritten character recognition tasks.

ABSTRACT

Convolutional networks almost always incorporate some form of spatial pooling, and very often it is alpha times alpha max-pooling with alpha=2. Max-pooling act on the hidden layers of the network, reducing their size by an integer multiplicative factor alpha. The amazing by-product of discarding 75% of your data is that you build into the network a degree of invariance with respect to translations and elastic distortions. However, if you simply alternate convolutional layers with max-pooling layers, performance is limited due to the rapid reduction in spatial size, and the disjoint nature of the pooling regions. We have formulated a fractional version of max-pooling where alpha is allowed to take non-integer values. Our version of max-pooling is stochastic as there are lots of different ways of constructing suitable pooling regions. We find that our form of fractional max-pooling reduces overfitting on a variety of datasets: for instance, we improve on the state-of-the art for CIFAR-100 without even using dropout.

Motivation & Objective

• To address the limitations of standard 2×2 max-pooling, which rapidly reduces spatial resolution and creates disjoint pooling regions that hinder generalization.
• To explore whether a gentler, non-integer pooling factor α (1 < α < 2) could improve feature hierarchy and model robustness.
• To investigate how stochastic, overlapping, and pseudorandom pooling region selection affects performance and generalization.
• To evaluate FMP's effectiveness in reducing overfitting and improving accuracy, especially in the absence of dropout or extensive data augmentation.

Proposed method

• FMP uses a non-integer pooling factor α (e.g., √2, ∛2) to reduce spatial dimensions by a fractional factor rather than halving them as in standard 2×2 pooling.
• Pooling regions are selected stochastically—either randomly or pseudorandomly—across feature maps, introducing controlled variability in the pooling operation.
• Regions are allowed to overlap, which improves performance compared to disjoint regions by preserving spatial continuity and feature context.
• The method uses max-pooling within each region but varies the region placement and size based on α, enabling multi-scale feature representation.
• Pseudorandom and overlapping pooling regions are found to yield better results than random or disjoint configurations, especially when combined with data augmentation.
• The approach is applied in deep CNN architectures with multiple FMP layers, replacing standard max-pooling layers to maintain spatial resolution longer.

Experimental results

Research questions

• RQ1 Can a non-integer pooling factor α (1 < α < 2) improve generalization in convolutional neural networks compared to standard 2×2 max-pooling?
• RQ2 Does stochastic, overlapping pooling region selection in FMP reduce overfitting and improve performance on small datasets like CIFAR-100 and MNIST?
• RQ3 How does FMP compare to data augmentation and dropout in encoding invariance to small distortions and translations?
• RQ4 Does the use of pseudorandom vs. random region selection affect model performance, especially when combined with other regularization techniques?
• RQ5 Can FMP achieve state-of-the-art performance without relying on dropout or extensive data augmentation?

Key findings

• On CIFAR-100, FMP with α = √2 achieved a test error of 26.39% with 12 random weight initializations, outperforming standard 2×2 max-pooling without dropout.
• For MNIST, FMP with overlapping, pseudorandom regions reduced test error to 0.32% over 12 runs, significantly below the 0.54% error of standard 2×2 max-pooling.
• On the Assamese handwritten character dataset, the FMP network without data augmentation achieved 0.7% error (12 runs), outperforming the MP2 network with affine data augmentation (1.8% error).
• For CASIA-OLHWDB1.1, FMP reduced test error to 2.97% (12 runs), compared to 3.82% for a standard MP2-based model with dropout and data augmentation.
• On CIFAR-10, FMP achieved a 3.47% test error over 100 runs, outperforming a baseline MP2 model with the same augmentation scheme and matching the performance of a top Kaggle solution.