Competition AI vs human

Artificial intelligence has long since become part of our everyday lives. In many applications we are aware of the use of AI, in others we are not. Most areas of application for AI are in the area of supporting and relieving people in their professional and everyday lives. However, there are also AI systems that are ahead of us and surpass human abilities such as in these examples:

At the university KIT and its spin-off KITES in Germany, a team of scientists and researchers are currently working on a computer system that is able to follow and accurately reproduce everyday conversations within seconds – even better than humans and faster than other computer systems. AI especially has to overcome speech barriers like filler words, semantic breaks, stuttering and interruptions which are challenging for humans as well. Prof. Alex Waibel and his team already developed an automatic live translator, which is used to translate the lectures at the university from German or English into different foreign languages. Recognizing spontaneous speech is the most essential component in the process, since errors and delays in recognition makes the translation incomprehensible and therefore need to  be minimized. Currently, the human error rate is 5.5% and the system’s 5.0% within a second.

Moreover, for the first time, AI achieved better results in the Stanford-Test for reading and comprehension than a human  – the record was achieved by both the Alibaba and Microsoft AIs. The Stanford-Test, which determines whether a participant understood enough of the content of the text after reading and answering a series of specific questions about it. This is believed to be one of the biggest challenges for AI systems. As part of the test, the AI softwares read texts from a selection of over 500 Wikipedia articles and then answered specific questions like “How is rain formed?” as accurately as possible. The Alibaba AI reportedly scored 82.44 on the Stanford-Test, the Microsoft AI even managed to achieve 82.65 points. Both AI systems beat the human score of 82.3 points. In the future, this AI can be used for customer service applications, museum guides or to answer medical questions on online platforms.

Another helpful AI based software is the one from Deepmind. This software “Streams” is monitoring hospital patients and is moreover able to predict a worsening in a patient’s condition and enables doctors to intervene at an early stage. The AI system recognises risk factors that lead to deterioration. So far, the project only relates to kidney damage. The biggest advantage of the software is not its superiority in the specialist medical field though, but that computers are able to observe patients 24/7 and therefore are able to support and unburden medical staff.

Google researchers are currently working on a project in which AI software computes other AI software. The intelligent software AutoML was developed in order to program another artificial intelligence software. The software has in fact developed an AI system that surpasses the abilities of human developers –  it’s called Nasnet. Nasnet is an image analysis tool that focuses on recognising objects. In an image recognition test, it was able to categorize 82% of the images by their content. Another task was to mark the location of multiple objects in an image, which is a crucial and essential task for augmented reality and autonomous robots. Nasnet, the auto-generated system, was able to score 43%. In contrast to that, the best human-built system achieved a total score of 39%. This means that AI coded by an AI performed better than any AI developed by a human.