NISHANTH RAJ DHINAKARAN
We all know that technology has been rapidly progressing throughout this era, with brilliant and innovative designs dominating our world in nearly every aspect. However, a particular widely used sector in technology that continues to marvel us even to this day is Artificial Intelligence.
What many people don’t realise is that the integration of Artificial Intelligence is almost universal, from recommending shows on Netflix to your own personal Google Home/Alexa. But it won’t end there: this is just the beginning. Even if we’ve gone from simple AI programs that could beat you in a chess game to more intricate AI that can drive autonomous cars for you, we’ve still only just scratched the surface of what could be achieved with AI's power. In this article, I’m going to explain the basics of AI/Machine Learning, which I hope will captivate your interest if you plan on studying engineering or computer science in the future.
Although historically Alan Turing’s “Can machines think?” and McCarthy’s “Machines that could think autonomously” were considered the definitions of AI, such definitions have evolved significantly over time. Artificial Intelligence (AI) is considered to be the capability of a computer system to mimic a human’s intelligence, solving problems without being explicitly programmed to solve them. Through maths and logic, a computer can find reasons to make decisions. The intelligence is artificial because it is not human, but is created by humans for it to perform human activities.
Artificial Intelligence
There are 2 types of AI:
- Narrow AI (Weak)
- This type of AI operates within a limited context and is often focused on performing simple, specialised tasks to great ability. However, since such AIs are restricted from expanding in function, they are considered less intelligent than human beings.
- However, this does not mean that these AIs are not useful. In fact, they are the pinnacle of AI in our generation, in terms of the numerous breakthroughs in the last decade. Some examples include Google, Image Recognition software and Personal Assistants (Siri, Alexa etc).
- Artificial General Intelligence (Strong)
- AGI is pretty much what we see in movies.
- It has great intelligence: that of humans or potentially superior. It would be able to apply its intelligence to solve multiple problems. As of today, we haven’t achieved the perfection of strong AI, however, in the coming years, don’t be surprised if one hits the news.
If we look into a more structured aspect of AI, 3 qualities stand out:
- Intentionality - AI systems can make decisions from historical/real-time data.
- Intelligence - AI systems often use techniques such as Machine Learning and Deep Learning (more on this below) and data analytics to allow them to decide intelligently (to almost mimic the thought process of a human).
- Adaptability - AI systems can learn and adapt using the information they have and by making further decisions. As they learn over time, their decision-making capabilities improve, leading to better decision outcomes.
As you can see, most of these characteristics are very human-like. Humans can make decisions from the information they perceive and can learn from their mistakes in order to improve their decision-making skills. It’s just like if you get a question wrong in a maths test, and you practice and practice until you find an efficient method to get to the solution, so you can do well in your next test.
Machine Learning
Well, we have AI, so what is Machine Learning? Looking at the figure below, we can see that Machine Learning is a key subset of AI, alongside Deep Learning.
Whilst AI and Machine Learning may be closely related, they’re not the same. It is important to understand that Machine Learning is a subset of AI. Whilst an “intelligent” computer uses AI to think like a human, Machine Learning is more the technical process of how a computer system can develop such intelligence. It is an application of AI, using mathematical models (such as graphs and networks) to help a computer learn. Based on the experience of being exposed to data sets, the computer will be able to improve its decision-making skills. As Tom Mitchell from Carnegie Mellon University puts it:
A computer program is said to learn from experience E with respect to some task T and some performance measure P, if its performance on T, as measured by P, improves with experience E.
Within Machine Learning, there are different types of algorithms:
- Supervised Machine Learning - In quite simple terms, the objective here is to come up with a mapping function (f) that describes an input (x) and an output (y). In other words, this is your function, which could be linear, quadratic or any other form, that maps input features to an output prediction. The parameters (coefficients) of the function are set to make it as accurate as possible on a given dataset, and the calculated function can then be applied to new data to obtain similar sets of results.
- Classification algorithms are used for categorical values (qualitative usually).
- Regression algorithms are used for real values (more quantitative or numerical).
- Unsupervised Machine Learning - You aren’t given any outputs here, only a set of input data. The computer must figure out the relationships and patterns within the data itself.
- Clustering is when you discover groups in the input data (a certain number of people have been grouped away from another certain number of people due to a characteristic).
- Association is when you discover rules in the input data (certain types of people tend to follow a given rule).
- Reinforcement Learning is unlike the others. In RL, the computer is an "agent" in a scenario, who learns the best path (of choices made) to take to gain a maximum "reward". Machine Learning models utilising this type of learning are sequential, as a sequence of decisions is made. These models are effective for learning how to play games (see Atari RL results), as the computer must use trial and error to get to a solution for a problem.
Deep Learning
Finally, we have Deep Learning, which as you can see from the diagram in Figure 1, is a subset of Machine Learning. However, it is often thought to be the next generation of Machine Learning. This is where it gets more advanced because rather than just grouping data and calculating functions, Deep Learning can use multiple layers of calculations to extract more complex functions and "learn more" from the same raw inputs.
A brilliant example of this is in Image Recognition apps. Rather than recognising matrix pixels (lower levels), Deep Learning can understand more complex features through layers of algorithms. This could be the details of your eyes, nose, mouth and how they are shaped or what tone/colour they have, and so on. Since Deep Learning can understand data from a lower level and process it to higher levels, it can improve its performance over time. Deep Learning is heavily associated with the structure and function of the human brain, which is why it is often dubbed “artificial neural networks”, as it tries to mimic the human learning process.
A common and ground-breaking example of Deep Learning being used is in Tesla’s autonomous cars. Tesla's AI team has developed a powerful chip that holds complex Machine Learning algorithms designed for self-driving. They have produced a schematic diagram visualisation in the form of complex webs (almost like a human brain), which reflects the structure of Deep Learning models:
That Deep Learning is so advanced can be explained by the fact that it uses both supervised and unsupervised learning. Not only that, but its employment in natural language processing, audio recognition (in your personal assistants) and medical image analysis (at hospitals for X-rays) is really state of the art in the world today.
Because it is so advanced, Deep Learning exceeds older learning algorithms in terms of performance. As Professor Andrew Ng states in his course on Deep Learning, Deep Learning is so good because it is extremely proficient in applying supervised learning in different situations. One important property of the Deep Learning neural networks is that the results of Deep Learning get significantly better with more data, more models, and more computation, while other models might stagnate.
Concluding this article, I hope I have given you a much better understanding of the different types of AI and how important they are in being able to solve repetitive, data intensive and superhuman tasks, conclusively saving time and effort. As AI advances in the world today, we as human beings must not fear its ever growing power but harness it in a sustainable and useful way so that we are able to bring efficiency into our lives. Whilst progress is being made (such as by DeepMind’s AlphaGo), we are still a few decades away from creating something that could exceed humans in all tasks. The demand for AI engineers just keeps on growing, which offers invaluable opportunities to be a part of something influential, and it’s up to you whether you would like to be a part of this experience.
Editor's Note: Nishanth would like to echo a recommendation an author has previously made on WBGS Looking Glass for Andrew Ng’s Coursera course which teaches how data is grouped, modelled and developed to enable a machine to learn. It is one of the best resources on the topic available to date. This is the source for much of the content in the article.