OK – the title might be overstating it a bit, but I’m a test analyst, and we get tired of hearing about how “automated testing” replaces the need for manual testers.
I recently saw this TED talk about machine deep learning.
I was somewhat surprised about how much the process for teaching machines was so akin to one of the fastest ways to teach expertise to humans – the exposure to 200-300 excellent examples in a short time, with a small amount of expert guidance, and seen in this presentation by Kathy Sierra.
I was thinking, it might not be so long after all, before machines are doing all our observational jobs, if they can learn pattern recognition like that (and they are more accurate and much faster and never forget…)
And then I saw this wonderful talk about aliens (or maybe not aliens). Even though number crunching is one of the primary uses for computers they could still miss evidence that humans could pick up, and more significant than that, they don’t see what we don’t program them to see. That is, we can see things that are not expected, but because we don’t expect an anomaly, we don’t make our computers see it. The pattern recognition a computer uses might throw a wobbly when confronted with a 3-wheeled car, or a wrecked car, but we would still see it is a car. Computers look for recognisable or definable characteristics. Humans characteristically have 2 arms, but we recognise that there are exceptions… We might be able to teach a computer the pattern recognition to sex chicks, but it won’t cope with a hermaphroditic or mutated example (unless we thought of that possibility ourselves, first).
Humans have observational skills that cannot yet even be explained, like the amateur astronomer that Bill Bryson talks about in “A Short History of Nearly Everything”. Reverend Robert Evans can see supernovas…
To understand what a feat this is, imagine a standard dining room table covered in a black tablecloth and someone throwing a handful of salt across it. The scattered grains can be thought of as a galaxy. Now imagine fifteen hundred more tables like the first one—enough to fill a Wal-Mart parking lot, say, or to make a single line two miles long—each with a random array of salt across it. Now add one grain of salt to any table and let Bob Evans walk among them. At a glance he will spot it. That grain of salt is the supernova
We may not all be exceptional, like this, but these examples show there is an argument yet to justify the inclusion of human observational skills on top of machine pattern recognition and deep learning even when the machines get good at it and are in common use. We learn new things when someone sees something they don’t expect, and says “WTF?” and follows the lead. Test wisdom says “the difference between what we know and what we need to know is why we test in the first place”. Machines might be able to take care of the things we know, but it will need a true AI before the need for complementary human analysis can be questioned.