What Neural Networks Teach Us About Schizophrenia

Pretrained Artificial Neural Networks used to work like a Blackbox: You hand them an input and they predict an output with a certain probability -- but without us knowing the internal processes of how they came up with their prediction. A Neural Network to recognize images usually consists of around 20 neuron layers, trained with millions of images to tweak the network parameters to give high quality classifications. The layers consist of neurons that are trained to only forward information if they recognize one specific image feature, resulting in an action potential that serves as an input for the neurons of the next deeper layer. Each layer gets the information of the previous layer and supplies information to the next one until the output layer states the networks prediction. How many neurons of a certain layer fired their action potential implies how strongly the layer recognized its training features in the provided image. One of the little things we know about the functionality of Neural Network that recognize images is that each additional layer extracts higher level features of the image: While the first layer looks for edges and corners, the middle layers recognize shapes, the last layers whole objects and compositions.