Page 68 - Monocle Quarterly Journal Vol 3 Issue 2 Spring

MONOCLE QUARTERLY JOURNAL | DEEP LEARNING
3.6 CONCLUSION
What we know for sure is that artificial intelligence is a powerful tool in mankind’s never-ending quest to become more dominant as a species. Whether we should respect or fear artificial intelligence’s future potential is less certain. In the last decade, AI has made great strides in its practical applications, infiltrating almost every sector and industry in some way or another. And yet, despite the many new and exciting uses that AI has achieved, it may seem somewhat surprising that even the most “intelligent” machines still cannot perform functions that we as humans take for granted every day.
Whilst AI has become very good at performing certain tasks – such as recognising images for example – other natural human endowments have largely escaped the grasp of machines. One such barrier to achieving what we deem true intelligence is the difficulty for machines to grasp perhaps our most human of all traits – the natural acquisition of language and the multitude of social cues that are intertwined with conveying a specific message from one person to another. This failure of artificial intelligence to truly understand the nuances and contexts of seemingly simple lingual utterances has been made painfully obvious in a number of highly-publicised and extremely embarrassing instances. The most obvious and distressing of these examples is perhaps the case of Microsoft’s Tay.ai Twitter chatbot, which after only half a day online turned from a sweet and chatty teenage girl – as per her programming – into a Nazi-loving misogynist, thanks to the nefarious influences she encountered whilst talking to her fellow Twitter users.
Despite AI’s linguistic shortcomings, the successes achieved in the pursuit of computer replication of the capability of human vision have been no small feat. As one of the earliest pursuits of artificial intelligence research – dating back to the creation of the Perceptron in 1957
– AI has thrived in the domain of image recognition, to the extent that machines are now in some respects better at this task than humans. This fact has been proven since 2015, when in the annual ImageNet Challenge the winning team’s application more accurately classified images into categories than the average person. By 2017,
Whilst AI has become very
good at performing certain tasks – such as recognising images
for example – other natural human endowments have
largely escaped the grasp of machines.
the level of human image classification was exceeded by 29 of the 38 competing teams, all of which achieved an accuracy of over 95% for putting images into one of the thousands of prescribed categories. And in 2018, given the success of these now almost-commonplace applications, the difficulty of the challenge will be increased greatly, by building up a database of 3D images that must be recognised by competitors’ computer vision programs.
Computer vision – as perhaps the most closely- human trait artificial intelligence has achieved thus far – has opened up a world of applicable possibilities. In medicine, for example, AI can today diagnose the easily- curable but potentially-blinding illness called diabetic retinopathy, as well as pneumonia, more accurately than doctors, by applying machine learning techniques to thousands of X-rays. In the increasingly competitive and well-funded industry of self-driving cars, computer
66