The next time you read a media report about artificial intelligence’s latest success, remember all the failures from which this is likely to have been plucked
Read any media article on artificial intelligence or ‘AI’ these days and you are likely to come away with the impression it won’t be so very long before your life is being run by the Matrix and your boss is a marginally more benign version of the Terminator. Apparently, however, there are still a couple of areas where robots do not yet excel and so, to even up the score a little, we are going to focus on those.
To be fair, neither area will change the world or offer you much of a new career path but, if they puncture the aura of invincibility around AI and perhaps help push the date of the robot apocalypse back a few weeks, then that is no bad thing. Both findings stem from experiments by research scientist Janelle Shane, who had thus far tried to ‘train’ her own neural network to create recipe lists and Pokémon characters – with mixed results.
Training a robot to flirt
Training a neural network essentially involves feeding it masses of raw data so it eventually starts to recognise and then repeat patterns. Shane did this with chat-up lines in the hope her network would start generating its own – although the initial results were not promising. “De’ver dincing you bangting Won Pantt ore”, for instance, may well be deeply romantic but sadly not in any language known to humanity.
Further training led to some results that at least resembled English sentences – “You must be a triangle? Cause you’re the only thing here” and “Hey baby, you’re to be a key? Because I can bear your toot” being a couple of examples – but the point is that these represented just a few out of thousands of lines. And yet, in most stories about AI, you only tend to hear about the successes.
Granted, among the dross were “Do you have a pence because you stole my heart bat?”, “I had a come to got your heart” and ‘You look like a thing and I love you” – and we have probably all overheard worse being slurred in a bar or club on a Friday night. Even so, you would have to chalk chat-up lines up as a failure for AI – at least for the time being.
50 shades of Dorkwood
Shane’s other experiment concerned paint colours and, this time, she fed her network 7,700 shades marketed by Sherwin-Williams (think ‘American Dulux’) along with their RGB values, which are the number sets that enable computers to define colour. Once again, the first attempt was unsuccessful, with the network churning out a selection of dull, all-but-identical shades of grey.
Boosting the AI’s creativity settings did lead to a broader palette of colours but the names it attributed to them – for example, “Bylgfoam” and “Dondarf” – were more appropriate to characters from Lord of the Rings. As with the chat-up lines, however, the third attempt was a bit more successful still and the network was able to recognise a few colours, including white, red and grey.
As with the chat-up lines, however, the network struggled with the demands of creativity and its proffered list of somewhat questionable names – such as “Dorkwood”, “Light of Blast”, “Grey Pubic”, “Sindis Poop”, “Stanky Bean” and “Turdly” – seem unlikely to be ones that will ever emerge from a Farrow & Ball brainstorming session. Strike two, we would suggest, for artificial intelligence.
As we say, it is not much and yet, when you read the breathless reports of every AI success, it is worth remembering all the failures from which these will have been plucked. The robot apocalypse may well be coming. But not today.