Amplifying The Uncanny
Analysing the methodology and applications of Machine Learning and Generative Adversarial Networks (GAN) framework.
Computational tools and techniques, such as machine learning and GAN, are definitive to the applications of such technology for a generative purpose. The paper explores the exploitations of these deep generative models in the production of artificial images of human faces (deepfakes) and in turn invert its “objective function” and turn the process of creating human likeness to that of human unlikeness. The author highlighted the concept of “The Uncanny Valley”, introduced by roboticist and researcher Masahiro Mori, which theorises the dip in feelings of familiarity or comfort when increasing human likeness of artificial forms reaches a certain point. Using the idea of “the uncanny”, Being foiled maximises human unlikelihood by programming the optimisation towards producing images based on what the machine predicts are fake.
Machine Learning uses the process of optimisation (the best outcome) to solve a pre-defined objective function. The algorithms used to process data produce parameters that categorise what can be generated (by the choice of function). In producing deepfakes through the GAN framework, the generator serve to produce random samples and the discriminator is optimised to classify real data as being real and generated data as being fake, where the generator is trained to fool the discriminator.
Being foiled uses the parameters generated by the discriminator which predicts signs that the image is fake to change the highly realistic samples produced by the generator. It reverses the process of generating likelihood to pin-point at which point do we cognitively recognise a human face to be unreal, which relates to a visceral feeling of dissonance (the uncanny valley). When the system generates abstraction, where images cannot be cognitively recognised, I would imagine that the feelings of discomfort dissipates. In a way, Being Foiled studies the “unexplainable” phenomena of human understanding and feelings.
As a study, I feel that the generative piece serves its purpose of introspective visual representations of uncanniness. However, the work should exist as more than “aesthetic outcomes” and the learning can be applied to various fields, such as AI and human robotics, that develop and explore human likeness and machines.
The Artificial Intelligence field is quite advanced in the development of intelligent technology and computers that mimic human behaviour and thinking, threading the fine line of what is living and what is machine. Considering the analogue forms of art, Hyperrealism saw artists and sculptors, such as Duane Hanson and Ron Mueck, recreating human forms in such detail that it is hard to differentiate which is real and unreal visually. When it comes to robotics and artificial intelligence, what defines it to be “human” is the responses that are produced by the human mind and body. By studying the data collected on “normal” human behaviour, the AI systems generate responses trained to be human-like. “The Uncanny Valley” explores the threshold of human tolerance for non-human forms, where imitation no longer feels like imitation, which is often referenced in the field. With Being Foiled, the point where uncanniness starts to develop visually can be tracked and the information can be used when developing these non-human forms.
Where “Being Foiled” can be applied
When I was in KTH in Stockholm, I was introduced and had the experience of using and interacting with an artifical intelligence robot developed by the university. Furhat (https://furhatrobotics.com/) is a “social robot with human-like expressions and advanced conversational artificial intelligence (AI) capabilities.” He/ She is able to communicate with us humans as we do with each other – by speaking, listening, showing emotions and maintaining eye contact. The computer interfaces combines a three-dimensional screen to project human-like faces, which can be swapped according to the robot’s identity and intended function. Furhat constantly monitors the faces (their position and expressions) of people in front of it, making it responsive to the environment or the people it is talking to.
Article on Furhat:
“The system seems to avoid slipping into uncanny valley territory by not trying to explicitly resemble the physical texture of a human face. Instead, it can offer an interesting simulacrum of a face that interacts in real-time with humans. This offers an interesting middle-ground between alien robot faces and clunky attempts to resemble human heads using latex and mechanical servos.”
When interacting with the Furhat humanoid, personally I did not experience any feelings of discomfort and it seemed to have escape the phenomenon of “the uncanny valley”. It even seemed friendly and have a personality.
It is interesting to think that a machine could have a “personality” and the concept of ‘the uncanny valley’ was brought up when I was learning about the system. What came to my mind was at which point of likeness to human intelligence would the system reach the uncanny valley (discomfort) beyond just our response to the visuals of human likeness. Can we use the machine learning technique that predicts what is fake or what is real on images (facial expressions) for actual human behaviours (which is connected to facial behaviour in the Furhat system)? -> how I would apply the algorithm/ technique explored in the paper
An interesting idea:
Projecting the “distorted” faces on the humanoid to explore the feelings of dissonance when interacting with the AI system
While generative art cleverly makes use of machine learning techniques to generate outcomes that serve objective functions, the produced outcomes are very introspective in nature. The outcomes should go beyond the aesthetic, where the concept can be applied in very interesting ways with artificial intelligence and what it means to be human.