Two faces of Deep Learning
C.P. Ravikumar
If you have tried the voice search that Google provides, you will be impressed by what it can do. All you have to do is to click on the speaker pin icon on the right had side of the Google search box. Perform the following searches: Search for a person, say, "Donald Trump." If you are adventurous, try "Donald Knuth," the famed computer scientist who wrote "The Art of Computer Programming." Try the name of a poet or writer who wrote in your native language, say "Bankim Chandra Chattopadhyaya" or "Da. Ra. Bendre" or "Subrahmanya Bharathi". Try to locate a famous song from a movie in your native tongue, whether it is Marathi or Telugu or Kannada. Did you find what you were looking for? The chances are that you are nodding your head in affirmation. In my own experiments, I was breezing through with ease. I did have some difficulty in "Knuth" - I then tried the pronunciation that I recalled Prof. Garry Miller used to use in his class. He would use Ka-nuth and not K-nuth or Nuth. When I said "Ka-nuth," the search engine opened the doors for me. To impress me further, the search engine ever read out a passage from Wikipedia. Although I am yet to use it, there is an application called Google Goggles which can identify famous landmarks, personalities and paintings.
What a long way we have come from character recognition! In traditional computer science, we teach students to do things like linear search, binary search, and graph search. We leave aside things like pattern recognition to other courses such as AI. Applications such as voice search and face recognition make use of concepts from machine learning and AI. Deep Learning is the new buzz word in computer science. It refers to the use of a number of layers of artificial neural networks, each of which can help recognize some feature in the input. How do we recognize objects when there is a medley of them in a box - marbles, rubber bands, safety pins, buttons and such. We know some key features of the object that we are looking for - it is round, it has four holes in the center, etc. Perhaps a good way to understand deep learning is to think of the game of 20 questions, where we are allowed to ask at most 20 questions to guess what the other player is thinking of.
Deep learning is finding its way into a number of areas, such as software written for driving cars and factory automation. There is also concern about the use of deep learning. My memory goes back to the days of Ronald Reagan who started the Star Wars defence initiative. Those were the pre-cold-war days. Prof. David Parnas was leading the software development for this system resigned and argued that there was no way the software could be made zero-defect. Today, in contrast, we hear of software that can look at the scene ahead of the car, identify where the road is and steer the vehicle on the road as it twists and turns along a highway. What if the software incorrectly classifies a road?
We have already heard of hardware failures that have resulted in legal battles - a car that accelerated without the driver's intervention, a mobile phone that went off in flames due to overheating. Software bugs may lead to accidents as well. Law is about to become even more complicated with robots and AI-powered vehicles.
This is the topic of my article which has appeared in #Vijayakarnatakatoday. I have summarized it above for my non-Kannadiga readers.
