Thursday

In this talk I will introduce attention as a state of the art mechanism for classification of radio galaxies using convolutional neural networks. I will present multiple attention-based convolutional neural network models, that perform on par with previous state of the art classifiers whilst using 50% fewer parameters than the next smallest traditional CNN application in this field and producing attention maps. These attention maps provide a level of interpretability for otherwise potentially daunting deep learning models. By investigating different approaches to normalisation and aggregation methods, I will present local selections that can maximise the impact of the model as a whole. The resulting attention maps can be used to interpret the classifications made by the model and I demonstrate how the most salient regions identified by our attention model align well with the regions an expert human classifier would make use of for equivalent classifications. I will also demonstrate the value of the interpretability of attention, by using the attention maps as a diagnostic tool to investigate individually mis-classified sources in the labelled test set and broadly identify if the errors arose from the model or the data.