Industrialization was truly a remarkable development for human beings, factory productions and the idea of consumerism. With the help of machines, factories could now increase their output multifold and more number of people could afford products that were earlier out of their reach. Machines could now complete tasks that were earlier done by one or more human beings, and now do that in much shorter period of time. This fundamental prospect has driven technology over generations and continue to do so. We want machines to be able to do what humans can, and do that much more efficiently. And when you take that concept from imitating physical labour to now imitating mental labour, this is where the realm of Artificial Intelligence begins.
Artificial Intelligence evolves around empowering machines to think like humans and possess senses that humans have. And this is where various fields of artificial intelligence emerge from. Humans have eyes and can see the world through it. In artificial intelligence field, there are machines that can do the same. These machines belong to the section “Computer Vision”. An example of such machines would be facial recognition software. Humans can hear and understand sounds through their ears, the machines that can do that come under the section “Speech Recognition” with examples like Google Translator (audio) and Alexa. Humans can move around and interact with their environment. The branch of AI that develops such machines is called “Robotics”. Humans can read and understand languages, the AI machines that do this come under the section “Natural Language Processing” or NLP. Predictive text, autocomplete and auto-correct are common examples of this. We see here that the aim of AI is to build machines that can think by themselves and can perform human like activities.
Apart from our senses, AI also focuses on what humans can do using their brains. One of the most important characteristic of human brain is its advanced pattern recognition ability. This is a remarkable capability which has helped humans since the beginning of their existence when they looked up in the sky and recognized patterns in stars and used them as a guide for understanding seasons and also for navigation during their voyages. And when AI tries to mimic human brain capabilities, it gives birth to a section called Machine Learning.
At this point, we can divide Artificial Intelligence into two sub sections: Symbolic Learning and Machine Learning. Symbolic learning is the section where there is limited input and output expressed in form of symbols and if-then statements. It was the dominant paradigm of AI in the early days of research (1950s-80s).
Machine Learning is an application of artificial intelligence (AI) that provides systems the ability to automatically learn and improve from experience without being explicitly programmed. Think of the time when we learnt how to ride a bicycle. It was a series of repeated bouts of falling and pedalling. But with each repetition, our brain learnt a little more and slowly became adept at the task. Machine learning intends to do the same as it analyzes past data and improves its decision making in a continuous manner. Speech recognition and NLP are examples of machine learning implementation. In this, the system takes the input and statistically tries to match it to make a pattern like our brain does when we read a sentence. If we write a word, the machine will run mathematical models to match the word to a meaning by mapping its alphabets. This is called Statistical Learning. On the other hand, teaching a robot to ride a bicycle also involves the machine to learn the skill. But this learning will imitate the human brain much more deeply with continuous loop of feedback from which the machine learns and improves in the next attempt. This is called Deep Learning. Autonomous driving cars are an example of Deep Learning. The computer vision in the cars, by analyzing massive amount of data, begins to learn to identify a person from a lamp post, to understand traffic signals and stop signs. And it keeps improving as it continues learning, just like a human brain.
We now see the relationship between Artificial Intelligence, Machine Learning and Deep Learning. AI is the superset of ML which in turn is a superset of Deep Learning.
Machine Learning applications today do analysis with terabytes and even petabytes of data. The analysis can be broadly divided into two aspects:
Classification: This is the process of categorizing data into specific classes which can be treated as independent actionable entities. For eg: An ecommerce organization can feed its entire customer base data into the machine learning system and the system will classify the customers into different sections on the basis of age, gender, preferences, active login periods, inclination towards discounting, time spent on the website etc. Machine learning can analyze the data on more than 150 dimensions and provide in-depth classification.
Prediction: This is the process of generating output by the machine when applied on new data, on the basis of training that has been done on historical data. This could include predicting customer churn rate or stock price and trend prediction on the basis of historical data.
Machine learning models require a lot of data to get trained on. Deep Learning especially requires a huge dataset to train on before it can make accurate decisions. These models can also be segregated on the basis of their training:
Supervised learning: If you train the algorithm with the data that also contains the answers, then it is called Supervised learning. An example of this is the facial recognition software in Facebook that automatically suggests tagging your friends in a photo. The models already has images of your friends and it learns about their facial structure and finds them in the photos that you upload.
Unsupervised learning: If you train an algorithm with the data where you want the machine to figure out the pattern, it is called unsupervised data. An example would be machine learning models that NASA and ISRO would be using to find patterns in the unresolved part of the universe.
Reinforcement learning: If you give an algorithm a goal and expect the machine to achieve that goal through trial and error, it is called reinforcement learning. The example of a robot trying to learn how to ride a bicycle through trial and error is a good example of this kind of training.
Creating infrastructure to host various machine learning models, give them enough computing power to commence learning and to store the huge amount of data can be daunting and many users are turning towards public cloud where they can not only store petabytes of data at cheap prices but also take advantage of managed platforms for AI and ML modelling. Example of such solutions are Alibaba Cloud Platform for Artificial Intelligence (PAI), Cloud AI and AutoML in GCP and Sagemaker on AWS.