The researchers developed a new pooling method for convolutional neural networks to improve feature extraction. Their T-Max-Avg pooling layer can adaptively select important pixels from the input data, leading to better classification accuracy on different datasets compared to traditional pooling methods.