Analyzing decision polygons of DNN-based classification methods

Deep neural networks have shown impressive performance in various applications, including many pattern recognition problems. However, their working mechanisms have not been fully understood and adversarial examples indicate some fundamental problems with DNN-based classification methods. In this paper, we investigate the decision modeling mechanism of deep neural networks, which use the ReLU function. We derive some equations that show how each layer of deep neural networks expands the input dimension into higher dimensional spaces and generates numerous decision polygons. In this paper, we investigate the decision polygon formulations and present some examples that show interesting properties of DNN based classification methods.