“What an AI cannot create, an it does not understand”, states Ian Goodfellow – a top AI researcher currently working at Google. Presently he studies ‘generative models’ which are AI’s that can create or generate real world transmissions such as sounds, images, etc. He explains the intent of his research by saying, “If an AI can imagine the world in realistic detail—learn how to imagine realistic images and realistic sounds—this encourages the AI to learn about the structure of the world that actually exists”, ultimately, he means to have an AI that understand what it sees. To reach his goal, Goodfellow aims to make AI smarter through the utilization of AI. How? Through a concept or technique he classifies as “Generative Adversarial Networks” (GAN) or what we may call ‘dueling AI’. The idea is actually quite straightforward; imagine having two AI’s, where one is able to create sounds and/or images (creator) that are completely realistic –to the human eye- and, a second AI that evaluates (evaluator) if the sound and/or image generated is fake or not. Basically there exists a feedback loop, where the evaluator will initially identify a sound/image as fake, fostering the creator to learn and identify the parts that deemed its creation fake. Thus, allowing the AI to recreate the images/sounds resulting in outcomes considered real by the evaluator.
Essentially, what Goodfellow might have initiated is the evolution of Machine Learning (which we read about and discussed in the 2nd week of the course) from supervised learning to unsupervised learning. It occurred to me, is it possible that this is the initial push towards the independence of artificial intelligence in all its forms from us? And what will be the aftermath be?
Clearly in terms of Business it will increase service and customer satisfaction since it can increase the reliability of AI immensely. Maybe it could even overcome the issue of privacy and confidentiality. For example in healthcare an AI that incorporates GAN technology could construct ‘fake’ patient healthcare records that cannot be differentiated as fake (since they are so real to ones that exist) which then another AI could utilize to improve the diagnosis and treatment of patients.
To me its quite an incredible concept with immense applicability in real-life business situations, and can even deteriorate existing flaws that currently limit the use of AI – such as privacy. I would be intrigued to hear of your ideas on its application, or maybe potential pitfalls of it usefulness and reliability.
References:
https://www.wired.com/2017/04/googles-dueling-neural-networks-spar-get-smarter-no-humans-required/
https://www.forbes.com/sites/forbestechcouncil/2018/07/18/three-breakthroughs-that-will-disrupt-the-tech-world-in-2019/#67acadf31f87
