Artificial Intelligence is the buzz word right now and as somebody who is into data science, I have been experimenting with it quite a bit. These days this word is being used quite often and in many contexts, and given the potential of its application in many fields, I wanted to just step back and really understand what this term “AI” mean at an intuitive level. This is what I have gathered so far.
The common definition of Artificial intelligence (AI) is the intelligence displayed by computer systems similar to intelligent humans. So, what is human intelligence? Below are some characteristics that can help define intelligence:
Reasoning and problem solving
Ability to learn, improve and generalize
Perception, randomness and creativity
Now, lets understand the above in the context of a few tasks that a computer can be programmed to execute (of course, with varying degrees of success):
1. A game of chess: Computers have achieved a very high degree of success in chess. We all know in 1997, Deep Blue – a chess computer by IBM beat the world champion Gary Kasparov. Chess is a challenging game, yet it can be described by a set of moves and can be played by searching through all possible moves. The number of moves can really be large, but they can still be defined and that is how a computer with significant processing power can overcome human opponents. So, can chess playing computers be called AI? Yes, in very limited aspects of reasoning, problem solving and learning abilities.
2. Predicting whether a loan applicant is credit-worthy: In cases like this, computers have had a lot of success. Using machine learning algorithms and by providing many features (information), for e.g. in this case, income, debt levels, past credit behavior of the applicants, computers can be programmed to predict the chances that a loan applicant is going to default on his loan. In this process, the computer program is demonstrating signs of intelligence like reasoning, finding patterns and problem solving. As we train the computer on more data and evidence, it is going to learn from its mistakes and further improve the accuracy of its prediction. Additionally, it is also learning to generalize by generating high accuracy predictions for new loan applicants. So, as you can see, this AI is more evolved than the chess playing AI. In contrast with a formal, rules defined task like chess, this AI is achieving a task which is somewhat informal and uncertain.
3. Visually distinguishing between cats and dogs: Identifying images of dogs and cats is so easy for us that we donot even have to spend time thinking. That’s not so simple for computers. Humans have been seeing dogs and cats since childhood. In our brains, we have created a mechanism wherein we have defined a set of features associated with dogs and cats. These features are so well-entrenched in our subconscious that if someone asks what those features are, we may not be able to spell them out. In other words, since we have been looking at many cats and dogs since childhood, our brain has now the “experience” of distinguishing between cats and dogs. When we want to make a computer as visually intelligent, we must provide it with that experience, i.e. train it on a huge number of images of cats and dogs and provide features that’ll help it distinguish. However, we ourselves are not able to spell out the features differentiating cats and dogs in our minds, so, it is very difficult for us to create and provide such features to a computer. Deep learning algorithms, which is a special branch of machine learning, helps a computer program to create those abstract features on its own by looking through millions of images of cats and dogs. Thus, deep learning algorithms are making computers gain more human like intelligence involving vision, randomness and generalization.
Deep learning algorithms derive the features on their own by training on lots of data and this is where they differ from traditional machine learning algorithms, where features have to be pre-defined.
4. Self-driving cars: This is the latest area of AI where there is a lot of interest and research. What really differentiates these AI systems from the above is that they need to be much more evolved with ability to respond to unforeseen, random situations. A self-driving AI must understand the apparent left-right signal of the car in front, but also must interpret the not so apparent intentions of other drivers. For instance, if the car in front is slowing down, there could be some unforeseen hindrance ahead (accident, traffic jam etc.) and the AI system needs to prepare itself. This requires machines to have common-sense similar to how humans use common-sense to interact with the world. Of course, humans get this from experience, but to design algorithms with such common-sense experience will require training these algorithms on every possible driving scenario out there, which just is not possible. These are some serious questions that need to be addressed, as can be seen from some of the fatal crashes of autonomous cars. In May 2016, Tesla Model S crashed into a tractor trailer, because it didn't notice the white side of the trailer against the brightly lit sky. Deep learning models (after getting trained on a large number of scenarios), help self-driving AI’s to make sense of the physical world but they need to be combined with other approaches to deal with the uncertainty and randomness of driving in the real-world by infusing them with human level common-sense.
5. Writing a Poem: An AI that has the ability to do tasks like write a poem or a song surely has more evolved human like intelligence including language skills as well as creativity. Of course, the extent of creativity can be debatable, but the strides being made in this field at-least points to the beginning of such creative AI.
The below figure summarizes the points that have been made above:
Strong AI and Weak AI
All the above examples are those of weak AI where artificial intelligence has evolved in the context of a task. These AI systems have evolved to near human level intelligence in achieving particular tasks. However, we are still way off in developing strong AI – systems that can learn and apply in a variety of contexts, for example, learning to drive a car in a city road In USA and applying those learnings to drive a truck on the highway in China. Strong AI is like a human child that learns from different contexts and experiences, evolves into a system similar to a fully formed human mind and then can perform tasks without human intervention. Once that happens, there is a big possibility for the world of sci-fi movies to become a reality. Well! the day may not be far.
Launching soon !! a course on Deep-learning that'll cover the above in depth along with the algorithms and applications. Watch-out this space for more details.