A convolutional neural network reaches optimal sensitivity for detecting some, but not all, patterns

We investigate the performance of a convolutional neural network (CNN) at detecting a signal-known-exactly in Poisson noise. We compare the network performance with that of a Bayesian ideal observer (IO) that has the theoretical optimum in detection performance and a linear support vector machine (SVM). For several types of stimuli, including harmonics, faces, and certain regular patterns, the CNN performance asymptotes at the level of the IO. The SVM detection sensitivity is approximately 3-times lower. For other stimuli, including random patterns and certain cellular automata, the CNN sensitivity is significantly lower than that of the IO and the SVM. Finally, when the signal can appear in one of multiple locations, CNN sensitivity continues to match the ideal sensitivity.