A thought bubble

Google’s Machine Intelligence tags a photo

Thought Bubbles

In-flight security instructions are always boring, more so for frequent fliers. Not so if you fly with Air France. Air France security briefing videos are fun to watch, and hence more effective. Food served is also better than most other carriers. Good job Air France!

While everyone on-board had lunch and was ready to fall asleep, something reminded me of a recent event. Around July 2015, Google Photos had mistakenly tagged two black people as Gorillas. Google had quickly apologized for it and had also rectified the defect. Obviously this defect was not intentional; Google’s image recognition system had learned to identify people based on real-world data. And rectification might have been a long-term fix or some stop-gap patch. But this event definitely looked like a close call.

This made me think what if the tag was not an error, and just a reflection on some deep-rooted evolutionary process. Is it a coincidence that humans have evolved from apes, and a machine identifies the similarities in the two species? What if a machine tags me as a monkey in some photo; will it be a coincidence? Guess not. Evolution is no coincidence; rather a consequence.

Mental Ramblings

Above thought engrossed my brain for two full minutes. For next half an hour, I kept rambling over this thought.

It must be interesting how a machine could identify objects and people in images. How does it work? Google hires experts who have PhD in various Sciences. I am no match to a Doctor of Philosophy. But, what if I can go deep in a topic which is exciting to me. If a topic is more exciting to me, than any researcher at Google, I might go deeper than anyone else at Google.

Being positive is good, but getting swayed only on sentimental value is not wise. So I should focus on a concrete task. I’ll let my mind ramble a bit longer to be able to reach some conclusion.

Google’s Photo tagging system must have taken a long time. How exciting it would be, if a machine could learn a lot in very short time. How about if I can prototype such a system that learns very fast. Like months’ worth of learning achieved by current systems could be achieved in a few minutes? Like Trinity learned to fly a bird in the blink of an eye. Learning at very fast rate rate would make it useful for numerous interesting use cases. Great.

Then I had a slice of dry cake, and the sugar made me feel sleepy.