The Imitation Game

A.I. or artificial intelligence is here and it is here to stay. Defying the doomsdayers' "the famous skynet representation", enterprises far and wide are today in a race to prove that they can create A.I. based platforms or solutions that generate revenues and attract the best of talents. Despite some of the popular belief being propounded by some of the best minds of our generation like Elon Musk and Bill Gates of an impending A.I. doom, the doom is far far away and the doom may not be over the horizon itself.

What then is the future may someone ask ? A.I. or artificial intelligence as we see today, is greatly influenced by algorithms , some of which were written and conceived as far back as 1930s. Majority of these algorithms have been statistical in nature, which saw an upscale during 1990s. Algorithms like Naive Bayes using probability and algorithms of linear regression , SVD (single value decomposition) which made machines look intelligent by giving them an ability to predict based on a given set of data.

Little do people realize that these algorithms exploit the basic nature of machines, an ability to compute faster , on hundreds and millions of linear algebraic elements, typically matrices and vectors which essentially meant allowing the machine to look at the data, form patterns in the form of "y =Wx+b" (where y is the output,W is the weight assigned to an input , x and b is the bias) . The machine had to observe features x, learn the best match for W and b and predict y, intelligence right ??

A.I. transcended this statistical scene with the advent of today's coolest bit , the neural network (which presumably had been proposed quite a while back but the machines lacked the computing power to draw any inference from data). Today's neural networks are hugely responsible for the hype around A.I. and also responsible for the doomsdayer's prophecies.

The situation is further propounded by some of the coolest techniques applied in neural networks, namely recursive neural networks (RNN), LSTMs and reinforcement learning (a new aspect where Google's deep mind team trained a machine to play simple Atari games and even the difficult game like Alpha Go , defeating the best human players in technique and tactics). The reason for the hype ?? and you would laugh when I say this that we still don't understand how these networks work internally. We know how to build a network, how to train it but how does the network vibrate to adjust itself is unknown

Question we have to ask is whether this is true A.I. or A.G.I (artificial general intelligence) and the answer is still "No" . Pow !! take that doomsdayers !!! we are still not there which begs the question as to what is A.G.I and how would we achieve this in the real world .

Now, with these thoughts let's seg-way to the theory I am proposing hitherto and the story has it's roots in the famous movie "Imitation game"

The Principle of O.R.T (Observe , Refine and Fine Tune)

For those of you who have watched the Imitation Game know this and for those you who haven't , my recommendation is to go and watch this. The movie doesn't o justice to the true brilliance of Alan Turing but nonetheless leaves the watcher with this impending thought of what was Turing envisioning. The story abstract from Wikipedia is as follows

"In 1939, newly created British intelligence agency MI6 recruits Cambridge mathematics alumnus Alan Turing to crack Nazi codes, including Enigma -- which crypt-analysts had thought unbreakable. Turing's team, including Joan Clarke , analyze Enigma messages while he builds a machine to decipher them. Turing and team finally succeed and become heroes"

Taking Turing's thought forward and applying it to a scale we have with software today: True A.I. can only be achieved via imitation , imitation of human life, imitation of the myriads of sensory motor inputs and outputs which the machines called homo sapiens do day in and day out

Someone said "Imitation is the best form of flattery" and I truly subscribe to that idea . For AI to succeed in the enterprise world , it has to imitate the human as humans have done all their lives . We are born with no knowledge of the world but we imitate by looking at people . The next generation AI has to follow the same principle.

The premonition is also based on how I viewed my own toddler . Little ones are born with no knowledge of the world but they try and imitate it like a sponge . I had always believed that if a machine grew with a toddler and observed the patterns he/she did, a machine would attain intelligence as a toddler gains. I found this evidence with my little one when I observed him follow sounds and comprehensions. he not only did started recognizing objects but also imitating me and my wife. The comprehension extended to people around him with chuckles and click of tongue as and when he observed people and behavior

Voila, I had my answer and I have been contemplating on this aspect and have been researching on a theory to be applied in practical enterprise world.

The theory is playing the "Imitation game" . By applying techniques like reinforcement learning or deep neural nets, machines need to observe and replicate and then fine tune or what I like to call as the ORT principle .

Once a certain pattern of behavior is learnt , the machine becomes a personalized cookie cutter for a job "I" do . This achieves hyper personalization of machine in the "me" space. Post this learning, machines need to observe other machines and then get apply the same principle of ORT in a cyclic fashion..... till infinity . These become generalized thinking and observing machines the way a child does.

The ORT principle would have neural networks playing the part of Observe, Refine and Fine Tune. Observation would allow neural networks to collect as much data from different inputs of the machine, its vision, sounds and even touch. Refining would include refining it's own hyper personalized knowledge and fine tuning would include optimization of this applied technique to the data learnt so far. You might be thinking this would need super computers , isn't it ?? The answer is both yes and no .With the advent of cloud and distributed infrastructures, these can be different elements of the same machine in a distributed approach.

Applying O.R.T. would become true AI or AGI as we call it .

These are random thoughts. Please feel free to comment and provide your comments and suggestions as well