Machine learning specialists at MIT have created Norman, a neural network inspired by Hitchcock's split personality psychopath Norman Bates, a character in the novel by Robert Bloch and Alfred Hitchcock's film of the same name, Psycho. Fortunately, the only function of a neural network is to analyze Rorschach spots. The developers argue that the project's sole purpose is to demonstrate how training sampling can affect artificial intelligence.
The Rorschach test itself was invented at the beginning of the last century by the Swiss psychiatrist Hermann Rorschach. Its essence lies in viewing ten images, which show symmetrical blots. The patient undergoing the test will be forgiven to interpret each of them by talking about the first association that came to mind. Rorschach and his followers believed that with the help of this test it is possible to determine the mental state of a person, since everyone who suffers from this or that mental disorder reacts differently to the shapes and colors of blots and sees completely different shapes and images in them. The real effectiveness of this method has not yet been confirmed, however, this does not prevent it from being used in psychiatric practice around the world.
Pinar Yanardag and his colleagues at the Massachusetts Institute of Technology also adopted the Rorschach test, which became the basis for a new neural network that was created using deep learning methods. As a base for a thematic sample, the researchers used descriptions of images from the thematic community on Reddit about talking about death: for ethical reasons, the developers do not name it.
Based on the collected data, the neural network was trained to analyze and interpret Rorschach spots. The responses of the neural network were then compared with the responses produced by the neural network trained on the MsCOCO dataset assembled for image recognition. Interestingly, the results were very different. A neural network trained on a standard dataset saw a "vase of flowers" in the spots, and Norman saw a "shot man".
Norman is a fairly simple neural network, and it has few functions. Its only practical application can be computer diagnostics of the responses of real people to the Rorschach test. In fact, the developers wanted to show how artificial intelligence can be biased depending on the training set.
Nikolay Khizhnyak