This is the year where we have started working on the self-driving vehicles and will be ready to drive these cars in a couple of years. The artificial intelligence has reached a level where there is no looking back. In spite of coming such a long way, this technology still makes mistakes and fails badly in doing its work.
For humans which is a Panda, could be a gibbon for a machine, or a school bus for you could be an ostrich for AI. Researchers from France and Switzerland showed how adversarial perturbations could cause a computer to mistake a squirrel for a grey fox, or a coffee pot for a macaw.
How this all happens?
Think of a child who starts learning numbers for the first time. When a child learns something new, she recognizes it on the basis of certain characteristics. For example- one is tall and slender, six and nine have one big loop and so on. Once recognizes, they pick up those digits easily.
Machine learning algorithms learn to read the world in a similar process. Scientists feed a computer with hundreds or thousands of examples of whatever the computer needs to detect. The machines identify data as “this is a number, this is not, this is a number, this is not.” It picks up the characteristics that humans have feed in it.
But, unlike a human child, the computer doesn’t pay attention to details like a cat’s furry ears or the number four’s distinctive angular shape. It only speaks what is recorded.
Instead, it’s only looking at the individual pixels of the picture. It doesn’t even check all the pixels. In case, if the majority of number ones have a black pixel in one particular spot and a couple of white pixels in another particular spot, then the machine is likely to make a call after checking that handful of pixels.
The only way to completely avoid this is to have a perfect model. Machine learning algorithms are usually scored by their accuracy. Another solution for this could be to put the programmes through their paces. Create your own examples of what a particular machine would do during perturbation and show it to the machine learning algorithms.