Pooling in neural networks is a way to combine features and make computations easier. Traditional pooling methods like max pooling are fixed, but a new method using recurrent neural networks (RNN) can learn and adapt to data. This new approach, called a fully trainable network (FTN), improves network performance by customizing pooling to the data and other network parts. Tests show that RNN-based pooling can match existing methods and boost network performance, especially on small networks. For example, on the CIFAR-10 dataset, FTN reduced error rates by seven percentage points compared to regular CNNs.