Artificial intelligence , Machine Learning and Deep Learning 101– Clearing the basics first

We find often the terms Artificial Intelligence (AI) and Machine Learning(ML) being used interchangeably. Popular movies like Matrix demonstrated a lot of AI and the differences between machine learning, AI and deep learning can often get unclear.

AI is a much broader concept where we are expecting a machine to replicate some tasks commonly associated with human intelligence. This can include planning, understanding language, recognizing objects and sounds, learning, and problem solving. Machines can perform intelligent tasks but not certainly limited to a single repetitive one – they can actually adapt and learn and do more.

Machine Learning on the other hand is a branch of AI but in more specific context.

Machine learning is an application of AI that provides systems the ability to automatically learn and improve from experience without being explicitly programmed. It can be seen as one of the ways to achieve AI.

Can we achieve AI without Machine Learning? Yes, but that means we are writing billions of programs for every scenario and rules and decision trees.

AI can be a ton of if-then statements - rules explicitly programmed by a human hand. Approaches like these are called Rule Engines, Expert systems or Knowledge Graphs. Simple put these are very early forms of AI.

Alternatively, ML is a way of training the program instead. The algorithm is fed data and then it adapts, adjusts and trains itself.

AI can be categorized as General and Narrow. Programs that mimic human intelligence in total is General while if they only do some functions like recognizing images it would be proper to be called narrow.

Methods of Learning

There are three broad methods of Learning. Supervised and Unsupervised.

Supervised – simply put this is where you have a teacher.

We teach or train the machine using data which is well labelled meaning some data is already pointed out as a correct answer. When new data then is presented to the machine, the machine (supervised learning algorithm) them analyses given data and produces an outcome.

Let’s say there is a basket filled with different kinds of fruits. Next, we tell the algorithm certain features and labels. Say if its red and flattened on top and bottom, it’s an apple. If its round and green and small, a grape. Now when a new data (grape) is presented, it will analyse and produce the output.

Supervised learning can be done by classification (category values like red, green) or Regression (real value like rupees or weight).

Unsupervised learning – no teacher (no help with features or labels). It is not told to the algorithm that features such as red and flattened at top means it is an apple. So, the algorithm has to act on information without any guidance. Usually the algorithm will then work on grouping the information based on similarities, patterns and differences.

Unsupervised learning is done by Clustering – find out inherent groupings in the data such as grouping customers by purchasing patterns. It can also be done by association like people who buy bread also buy butter.

Reinforced learning – learns by action and reward. Say every time a baby cries, there is a candy. Over time the baby will possibly be crying when it wants a candy...!!

Real world examples of Supervised and Unsupervised learning

Example of supervised learning – predicting house price based on input data. An example of unsupervised learning is handwriting recognition.

Supervised learning involves prediction. Say we want to find out which students will do well in the Entrance examination by analysing data on students who have already taken the test.

Think recommendations at Facebook, Amazon or Netflix.

Unsupervised learning explores and groups data – it may tell us subgroups of students taking the test. Okay ! a more real world example is say you have never watched a game of soccer. Now you land up watching one, and then immediately your brain starts grouping players by their jersey color etc.

Let’s jump into future – Deep Learning

Talks about AI began in early 1950s. Somewhere around 1980s machine learning was introduced and is increasingly popular today. Now we are moving into more complex structures that mimic the human brain – Deep Learning. Deep Learning is really machine learning put on steroids. An example is neural networks.

It’s a part of a larger machine learning methods based on data representations rather than task specific algorithms. Example uses of Neural networks are speech recognition, natural language processing, drug design etc.

What IS and NOT  Artificial Intelligence

In 1950, Alan Turing came up with the Turing Test – if you ask a machine some questions and at end if you cannot distinguish if the responses came from a machine or a human, the machine would have passed the test. Peter Norwig at Google though does not spend too much time on passing the test. He jokingly mentions that’s like “making a flying machine that can fly like a pigeon and can confuse other pigeons”. Instead he wants to focus on the principles of intelligence.

Some say the moment a machine does something intelligent, it ceases to be so. A machine beating a chess champion can at best be said as bright.

Some critics say the current level of AI can just be said as accelerated statistical modelling.

 So, a better definition as per Douglas Hofstadter: "AI is whatever hasn't been done yet." So, one criterion of AI could be machines that find solutions in problems they haven’t seen before- novel scenarios.. That’s what we humans are good at - isn’t it?