Designing a Synthetic Neural Network Layer for Image Classification (2019 - Present)

We designed a novel artificial neural network layer inspired by the biology and the mathematics of the human visual system to make image classification models more accurate and train faster. The first layer of a neural network serves as a filter to extract the most important features in an image. Constructed weights were used in the first layer and were fixed so that the same features were extracted each time. This design allows for more feature-specific neural network design and improves upon the performance of neural network training.

We also assessed the behavior of this synthetic layer against adversarial images, as well as that of models that take into account pose (the orientation and location of certain features of an image in relation to each other) including CapsNet.