Like flying cars and space colonies, for a long time Artificial Intelligence (AI) looked like just another extravagant futuristic vision from the 60s: great for Hollywood movies, but far from becoming a tangible reality. This has changed dramatically over the past few years, mainly thanks to the rapid progress in applied technologies such as deep learning and neural networks.
The potential impact of these technologies was spectacularly demonstrated between 2015 and 2017 by AlphaGo, an AI created by Google’s DeepMind Company and the first program to ever beat a human player at Go, a traditional Chinese board game famous for its high level of complexity. The victory of AlphaGo and its implications featured in an interesting documentary, and it drew the attention of the broader public to the consequences of the latest wave of automation. As awareness about the possible loss of jobs in professional fields such as medicine, law or programming is slowly gaining ground, we took a look at five highly specialized activities in which machines are already outperforming their human competitors.
Medicine: Researchers at an Oxford Hospital have developed an AI that can use heart and lung scans to diagnose deadly diseases. While this has also been routinely done by human doctors, computer programs promise to yield clearly better results than even the best medical professionals, as it is estimated that at least one in five patients are misdiagnosed. Earlier, more accurate prevention and fewer unnecessary operations could lead to enormous cost reductions.
Law: LawGeex, a start-up specializing in AI-driven contract review, pitted its three-year-old algorithm against 20 experienced human lawyers. The results clearly demonstrated the AI’s superiority, with an accuracy score of 94% (vs. 85% for the human lawyers) and an average time required to review the contract of 26 seconds (vs. 92 minutes needed by human competitors).
Art authentication: And surely not one that machines can handle. But this may be about to change, as demonstrated in a study by Rutgers University. A deep recurrent neural network (RNN) was fed with data from 300 line-drawings by painters like Picasso, Modigliani and Matisse, which it then broke down into 80 000 individual brushstrokes. While the analysis only succeeded when the brushstrokes were highly visible, the results were still stunning.
Programming: It seems like AI may also be our best choice when trying to design… another AI. At least that’s what four of Google Brain’s researchers discovered while trying to program an artificial intelligence that could accurately identify objects in real time. Their solution was to create AutoML,a machine learning model designed to create other machine learning models using an approach called reinforcement learning. The resulting AI surpassed anything designed by human programmers.
Reading: Another presumed milestone was reached in January this year, as teams from tech giants Alibaba and Microsoft independently created programs capable of beating humans at the Stanford Question Answering Database (SQuAD), a standardized reading comprehension test (actually a database, as the name suggests) developed by Stanford University computer scientists. However, as this article shows, headlines about the machines’ reading prowess were not completely accurate, leaving hope for humans – at least in this specific field.