Why machine intelligence sometimes gives crazy answers

Machine intelligence has developed considerably in recent years. However, sometimes even the most sophisticated platforms give nonsensical answers. A research team have been delving into the ‘black box’ of AI to find out why.

Technologists from the University of Maryland have been exploring machine learning algorithms to work out how they function and with a view to addressing errors. Previous approaches to such investigations have aimed to 'break' the algorithms in ways that involved taking out key words from inputs, so that the wrong answer is produced.

Taking an alternative route, the computer scientists decided to reduce the inputs down to the minimum required to yield the correct answer. This demonstrated that the correct answer could be obtained with an input of less than three words.

As an example, the scientists presented an algorithm with a photo of a sunflower and asked the question: 'What color is the flower?' This generated the correct answer (in this case 'yellow.') The scientists then discovered found the correct answer of ‘yellow’ could be produced by asking the algorithm a single-word question: 'Flower?'

With a second and more complex case, the scientists used the prompt, "In 1899, John Jacob Astor IV invested $100,000 for Tesla to further develop and produce a new lighting system. Instead, Tesla used the money to fund his Colorado Springs experiments."

Following this the researchers questioned the algorithm with: "What did Tesla spend Astor's money on?" This led to the correct answer: "Colorado Springs experiments." However, they also found that reducing the input to the single word "did" also generated the same correct answer.

These types of studies showed that keeping algorithms simple not only saves time but avoided the issues that can arise from algorithms producing the wrong answer, which the researchers attribute to over-complexity. This is partly because most algorithms are compelled to provide an answer, even where there is insufficient or conflicting data.

For the simpler approach to work, the input word did not need to have an obvious connection to the answer. This reveals how some algorithms react to specific language in a way that is different to people.

It is hoped that the research will help computer scientists to create more effective algorithms, including machines that can recognize their own limitations.

The research was presented to the 2018 Conference on Empirical Methods in Natural Language Processing, which took place in Brussels, Belgium during November 2018.

ABOUT US

With 21 years in providing R&D services for leading companies worldwide, TMA Solutions (http://www.tmasolutions.com) has intensive experience and capabilities to bring your products and solutions to a new level by embracing new technologies (big data & analytics, AI, IoT, mobile, cloud, etc.).

TMA Innovation Center (TIC) is TMA R&D arm focusing on collaborating with universities and partners worldwide to foster innovation and build a new generation of products and solutions.

TIC creates an open environment fostering idea & knowledge sharing, R&D collaboration, creativity & innovation, technology transfer and incubation.