Hybrid Pooling As A Balanced Approach to Reduce False Positives in CNNs

Pooling operations are essential in Convolutional Neural Networks (CNNs) for reducing spatial dimensions while preserving key features. Max pooling captures strong activations but can amplify noise by always selecting the highest values, which may lead to suboptimal representations. To address this, we propose hybrid pooling methods that combine max and average pooling in structured or probabilistic ways. These methods aim to retain important features while improving generalization and reducing overfitting. We evaluate them on CIFAR-100 using ResNet-18 and VGG-16 architectures. Our hybrid-random pooling method consistently outperforms traditional max and mixed pooling, achieving up to 7% higher accuracy. These results demonstrate that simple, randomized pooling strategies can provide robust performance gains without adding complexity to the model, offering an efficient alternative to conventional pooling techniques.