The big game-changer in AI in the last 10 years is that machines have started to learn. This means that they don’t have to be explicitly programmed. Take, for instance, the machine DeepBlue which defeated the Chess grandmaster Gary Kasparov. Now, this machine was not intelligent in the sense that the programmers had tediously programmed all the board positions and actions to take explicitly into the program. As Chess is not a complicated game (having only 64 board positions) this was an easy task to do. Some years later, the grandmaster of ‘Go’ game, Lee Sedol, played against another machine called AlphaGo. ‘Go’ is a very complicated game compared to Chess. It’s not humanly possible to program all the board positions. So, the makers of ‘Go’ programmed the basic rules into AlphaGo. The machine started playing against itself and learned from the mistakes it made. In the beginning, it was playing like a toddler. Soon like a teenager, an adult, a master, and a grandmaster. It defeated Lee Sedol, and the rest is history.
So, what are we meaning when we say that machines have started learning? Say, you show 1 million images of cats to a machine so that it recognizes what a cat looks like. Most likely it will guess the next cat picture accurately. However, if you had by mistake shown some dog pictures while you were training the machine, the results would be unpredictable. Remember the good old adage ‘GIGO’, which means Garbage In Garbage Out. If your data is incorrect, the output also would be incorrect. Now, assume that the machine was fed good data. If you show it a cat image and call it ‘billi’ (In Hindi language ‘cat’ is called ‘billi’), it will automatically infer that a cat is ‘billi’ in the Hindi language. The key here is data. The more data that you feed the machine, the better the outcome would be. The bottom-line, however, is ‘good’ data. A lot of time that data scientist spend is in wrangling and munging data until it is clean.
Machines have learned to write poetry or Shakespearean novels. Just by studying the style of a personality, they can create wonders. For example, if you feed the machine with all paintings of Leonardo Da Vinci, it can render any painting in his style. Machines are also used in writing articles for newspapers and magazines. And you will be surprised and shocked to know that they do a better job than us. A branch of machine learning called deep learning is particularly popular. Deep learning uses a swarm of neurons (perceptrons / sigmoidal neurons) to understand the data as much as they can. While learning, it keeps adjusting its weights across the network, until it gets a good output. The basic goal is to optimize the output, given a goal. This neural network works like our brain. We also comprehend things in a parallel fashion, albeit at a low speed (25-100 Hz). The machine is simply trying to emulate our brain and the learning rates are sky-rocketing.
Experts say that 2029 is going to be the year of Singularity. One single machine will exceed the capacity of a human brain. That is good news. Machines will augment us in ways that we cannot envisage right now. Humans are error prone. We have our thoughts, emotions and energy (very often not aligned) which makes us act in the way that we do. We get bored. We need to sleep. We get angry. A machine does not have these limitations. Thoughts are similar to what goes on inside a machine. However, we will have to program emotions into a machine (the basic ones) and then let it learn from the events that happen around it. A lot of research is being done in the field of empathetic computing. Rather than we becoming like robots, it’s better if it’s the other way around. This will help the machines understand us and find out their place in a world ruled by humans. Doomsday will not happen if we provide the right ethical framework for machines. We are on the crux of sharing our world with a new species. Let’s welcome the future.
God Bless !