My understanding of towers in inception architecture and in tensorflow terminology is that they are part of a neural network model for which separate computation can happen on forward phase and gradient computation phase of back-propagation, independently. My present understanding is based on another datascience.stackoverflow post What is a tower?, tensorflow documentation at https://www.tensorflow.org/tutorials/deep_cnn and blog post on inception architecture https://pseudoprofound.wordpress.com/2016/08/28/notes-on-the-tensorflow-implementation-of-inception-v3/.
However, when I read Xception paper (https://arxiv.org/abs/1610.02357) in introduction section "The Inception Hypothesis", it talks about average pooling tower. Does this mean it is referring to branches within a inception module as towers, even though independent backpropagation computation will not be performed on these branches?