Although Artificial Intelligence (AI) has been around for some time, it is still common to get caught up in the buzz without truly understanding what it means. People think that AI is a robot that can do things a smart person would, knowing everything and being able to answer every question. This is what television and movies have led us to believe. Creating these ‘conscious’ machines is the goal of researchers and professionals but we are not close to that yet.
AI is classified into two groups. General AI is the concept explained above where machines can intelligently solve problems without human input. A machine with general AI capabilities would have cognitive abilities and interpret the environment around it. It would be able to process this information far quicker than any human could leading to these sci-fi ideas of superior beings. General AI currently is beyond our reach but as the volume of data in the world grows and computing power increases, we will get closer.
In the here and now, we are in a time of what is called narrow AI. A machine with narrow AI capabilities is one that operates from a predefined
set of rules. This could be a Netflix recommendation engine or a voice command system like Alexa. Both are examples of artificial i
ntelligence but are fed from criteria or training data to function.
A good example is a driverless car which although impressive, is still narrow AI as it is given a set of rules to operate by. Until a car can understand the environment and think for itself, this will always be the case.
Narrow AI applications are driven by two subsets of AI known as machine learning and deep learning. The best way of explaining the link between the three is that AI is an all-encompassing term, inside of which is machine learning and then within that we get more complex deep learning.
Machine learning is often described as a method for realising AI. A computer or machine is loaded with vast amounts of data which it will use to train itself. The data might be labelled initially to make things easier. For example, if we want a machine to recognise photos of cats, we may load it with thousands of images of cats and dogs, labelling them appropriately. The machine will take a new image and find the label it matches to learn whether it is a cat or not.
Machine learning is the process of enabling machines to learn through data. The predictions the machine makes from that data is what we know as AI. If we go back to the example of Alexa. Alexa receives a voice command, interprets that using an algorithm (known as natural language processing or NLP), matches the result against all existing data stored in the cloud to find the appropriate response and sends that back as a reply. Alexa gives the impression of being a cognitive machine but is far from it.
There are four common machine learning methods.
- Supervised learning
This method takes existing data and trains a model to work out how to classify a new piece of data. For example, it could hold data on the symptoms of diabetes and when it receives blood test results of a new patient, it is able to diagnose accurately from the data. It will classify the patient as having diabetes or not having diabetes.
- Unsupervised learning
Unlike, supervised learning, these models will attempt to classify data without any prior knowledge. The algorithms look to find patterns themselves and put data into groups. A common example is something like customer purchasing behaviours. The algorithm won’t have existing labels and will decide on its own how to classify the data, often known as clustering. Imagine going to a party where everybody is a stranger. Your mind will probably classify people based on age, gender or clothing. You don’t know them but have still worked out the classifications.
- Semi-supervised learning
As title suggests, this is a mix between supervised and unsupervised learning. In our data, some items are labelled but some are not. Where you have vast amounts of data this can be quite common. A semi-supervised model would have some labelled data to know that classification does exist. It is then trained on unsupervised data to define the boundaries of what it is looking at and potentially specify new classifications that the human did not specify when labelling.
- Reinforced learning
This application is about positive and negative rewards for certain behaviours. This will be a common method in robotics where machines learn to optimise behaviour from experiencing positive or negative results. For example, if a robot found a TV remote and decided to throw it, it would break and be a negative result. However, pressing a button turns the TV on a produces a positive result so it continues to do it. The robot will continue this process until finding the best possible result.
Whilst every AI-based project is unique as they all run from different datasets and rules, there are key algorithms that you will find in the library of virtually every Data Scientist.
Deep learning is a subset of machine learning. In a sense, it is an evolved version of machine learning methods. It is inspired by processing patterns of the human brain known as neural networks. Whereas machine learning techniques will take an input of data and learn from it, deep learning neural networks learn through their own data processing.
Unlike in machine learning where even in unsupervised methods a human still jumps in if the model gets too confused, deep learning algorithms decide for themselves whether a prediction is accurate.
It is difficult to ensure deep learning neural networks don’t come up with incorrect conclusions but when it works, it can get us a step closer to general AI.
AlphaGo by Google is one of the most popular cases of machine learning. Google trained a machine to learn the board game Go which requires a lot of intellect. Without being told what move to make, the machine learnt the rules itself and began to outperform humans. If the computer had of been fed rules through machine learning, this may not be hugely impressive but the fact it learnt how to win on its own is incredible.
There are many types of neural network that are used for different applications.
Narrow AI can be seen everywhere from GPS systems to Alexa and recommendation platforms. Machine learning and deep learning have benefitted from large investments at the start of the 21st century as consumers seek ways to become more efficient and have an easier life.
However, the ultimate goal is artificial general intelligence, a self-teaching system that can outperform humans across a wide range of disciplines. It is thought that this could be 30 years away or some say as long as a century but as computing power and data evolve, we will continue to see more amazing developments.