Confirmation Bias No.1

Deep Neural Network, commonly known to us as ‘deep learning’, is an artificial intelligence (AI) that performs given tasks such as regression analysis, classification, and prediction by stacking algorithms so called ‘artificial neural networks’ that mimic human neuron structures. In order for deep learning to work properly as humans intend, it is necessary to input a large amount of refined data and train it. Deep learning learns by self-analysis of the correlation between given data under human guidance, extracting meaningful features, and remembering the direction of reducing errors as much as possible through the process of verifying them. This is similar to the process in which a human child encounters various external stimuli under the guidance of parents and teachers in the development process and acquires an understanding of the world through mistakes.

Recently, the human learning environment is undergoing rapid changes. Active learning through reading, discussion, and direct experience is weakening its influence, and passive learning through watching media content through various smart devices is rapidly replacing it. The problem is that AI, which is used in recommendation algorithms to maximize service providers' profits, learns users’ propensity and continues to induce users to access only specific content tailored to them. Here, propensity encompasses all of political, social, economic, and personal interests based on information that a user unintentionally exposes to a service provider through the use of SNS or portal sites.

AI, which has learned the user's propensity, train humans in reverse through content recommendation algorithm and strengthens the existing worldview, values, and preconceptions through constant repetitive learning. Content that is contrary to the user's tendency is intentionally pushed to the back rank and concealed as if it did not exist from the beginning. As a result, each individual lives in a world of digital confirmation bias, where they only see what they want to see and believe in what they want to believe in, in the barriers of contents that excerpt only a small part of the reality which has diversity and multifacetedness in fact.

Over the past decade or so, anti-intellectual behavior, which has been seen by groups who consider themselves "collective intelligence" and a number of individuals who claim themselves to be "intellectuals," has confirmed that they are seriously buried in their own confirmation biases. Certain people that correspond to the biases became idols of each camp, not human beings with multifacetedness. And the idols of "our side" were sacrosanct, and the idols of "the other side" were thoroughly demonized. The details of this phenomenon are so vulgar that attempts to flexibly tolerate and explain it based on the old postmodern pluralism are considered worthless, followed by endless despair.

But the desperate attempt to talk about something that cannot be said would also be the soul of art. Art through the methodology of pinching, twisting, and turning the phenomenon upside down, proves its value by itself, apart from its practicality. And its value will be maximized when an appropriate medium is supported in line with the artistic theme that the artist wants to reveal.

Deep learning, an AI created by mimicking human neural network structures and learning mechanisms, has sufficient potential as a new media to deal with new phenomena, that is, the current situation in which humans are "learned" inversely by recommendation algorithms and fall into artificial confirmation biases. Convolutional Neural Network (CNN), an AI technology that has a particular strength in image recognition and classification among deep learning, is based on biological research that activates a specific optic nerve by focusing on analyzing patterns of images entering a specific area of vision.

The artist builds and trains CNN with data related to the objects of courtesy confirmation bias. Learning the images of each former president and the data related to them, which are representative objects of extreme confirmation bias, AI analyzes the correlation of these data on its own, extracts its features, abstracts them, and memorize them. The stored features may be reproduced in a form that can be visually identified by humans through activation maximization or so called, feature visualization algorithms.

During training, AI extracts features by comparing and analyzing various data, corrects itself in the direction of reducing the error between the resulting value and the actual data. Feature visualization activates the memorized feature map by twisting the learning mechanism of AI in reverse, stimulating the deep neural network in the opposite direction, rather than in the direction of decreasing error. This causes AI to distort the image presented 'as it wants to see'.

The distortion is gradual. At first, it decorates the screen with patterns that feel soft and sometimes beautiful, but as the bias intensifies, the image changes into a form like a monster that does not exist in the world. The end result of this process, which has gone through 500 steps, shows terrifying but simultaneously ecstatic abstract forms, such as hell, where corroborative bias prevailing in modern people is visually reproduced by AI.