The Age of AI: What exactly is AI?

The promise of artificial intelligence (AI) has been around since World War II but it is only in the last decade or so that technology and engineering have gradually caught up with expectations. In fact, AI now seems to be on a winning streak. But before worrying about being taken over by robots, let’s get a clear understanding about what AI is?

Having looked at the roller coaster history of how AI has reached its current platform of exponential growth, we delve into the various modern approaches to AI.

Beyond your web definition which describes AI as being the opposite of natural intelligence, it is more precisely the area of computer science that uses machines to mimic human reasoning and consequently appear to possess a cognitive element in the way they operate and solve problems.

Early incarnations of AI were merely a set of programmed rules such as “if this happens do this… otherwise do this or that”. There was little intelligence with such expert systems which were restricted in nature and could not adapt to situations they weren’t explicitly coded for. On the other hand, modern AI systems are good at learning and identifying subtle cues and patterns from a set of data and then applying this ‘knowledge’ to process new data related to that same problem. This is similar to how humans reason and react and is where the ‘intelligence’ aspect is derived.

We learn how to do many things by studying or through repetition, or by observing others. We practice speaking a language by slowly and often subconsciously absorbing the many nuances of its grammar. We do not think about grammar when we speak; at least no fluent speaker does. Similarly, we learn to drive by slowly tuning our mind and body to react to external cues, mostly visual, by controlling car speed and direction. We can recognise a pothole from a manhole cover and learn to anticipate where and when speeding up can actually avoid an accident. Very often, it is very difficult to articulate a ‘reason’ as to why we do what we do even though the outcome itself is often proof of the validity of our choices.

AI systems are reducible to a set of mathematical operations. Their output is not a result of a set of coded rules but is learnt through the application of mathematical processes to massive amounts of data. The ones building these models or algorithms may be able to understand and explain the general concept of how their program generates certain outputs, yet in most cases they themselves are incapable of explaining how their AI application would react to specific data.

This approach, known as Supervisory Learning, accounts for nearly 99%10 of the present AI economy. It works by presenting an AI system with a large amount of data that has already been labelled with whatever outputs the system is intended to recognise. The system will ‘learn’ from this data to associate patterns in the input with an output, equally well with previously unseen data. Reinforcement learning (RL)11 is a variant of such an approach whereby the machine does not need to be explicitly taught using large volumes of tagged data. Instead, an AI program using RL, known as an agent, learns from the consequences of its actions from past experiences and by random choices akin to trial and error12. This is like getting someone who has never seen a bike, to ride a bike when the only guidance given is ‘not to fall’. Such learning methods are the subject of intense interest and research and have had much publicised results. These are the platforms that can learn to play games, from Space Invaders to GO, just by being let loose on a game and allowing them to determine strategies by trial and error. Clearly, however, whilst practical for game play, it can be difficult to apply such a technique to teaching a machine how to diagnose cancer! Training a driving AI by starting from a dumb place and depending on trial and error can be dangerous and expensive. Conversely, the classic classroom approach of compiling large enough labelled and accurate datasets to reflect real life driving conditions can also be prohibitive in cost and time.

So why not apprentice a machine to a human?

This is the philosophy of a new area of research known as Imitation Learning. Apprenticeship has worked well with humans for millennia and there is no reason for not extending the technique to machines, whatever the algorithm. Staying with the driving scenario, take a car AI agent and put it in the driving seat so it will generate all the motions (steering, acceleration, deceleration etc.) that it determines are necessary to get from A to B. The machine will process inputs from its GPS, cameras, proximity detectors etc. to determine the best course of action and act accordingly, only it will not actually drive the car - a human mentor will be driving the car. What the AI will do is gradually learn to adapt to its environment by comparing its actions to that of the human being. This training can take place equally well on a school run, a road trip or on delivery routes, so that AI can be exposed to a wider sample of real life scenarios. We just need to make sure that the human driver chosen is better than a dumb machine to start with! This can make training safer, cheaper and potentially faster and could be the nearest thing to the transfer of human knowledge to a machine.

Another approach to AI requires no pre-existing knowledge about the data and is known as Unsupervised AI. The power of such algorithms lies in their ability to use raw collected data without significant data preparation and pre-labelling requirements. They can uncover patterns and data clusters with little human intervention and may have a stronger claim to the ‘intelligence’ part in AI. In reality, it is still ‘just’ mathematics. One may encounter such techniques in the processing of geo tagged weather data to isolate or identify specific climactic patterns or in the analysis of market research responses to identify clusters of similar responses.

By far the most popular AI algorithms are neural networks, specifically deep neural networks (DNN) and derivatives. Significant research investment fine-tuned these algorithms to the point where in the 2012 ImageNet13 competition, a convolutional DNN achieved a 10.8 percentage point improvement in image recognition in one go. Overnight, DNNs took the podium and are now synonymous with AI. Such networks are extremely adaptable yet more than any other algorithms operate as a black box. It is near impossible to explain or rationalise the results obtained from such networks even when proven empirically right in many fields. This lack of ‘explainability’ is often undesirable and real AI professionals understand when simpler algorithms make more sense. As with many things in life, different problems call for different solutions that are not necessarily neural networks.


[1] A. Ng, “Andrew Ng: Why AI Is the New Electricity | Stanford Graduate School of Business,” 2017.
[2] J. McCarthy, “Arthur Samuel : Pioneer in Machine Learning.”
[3] Sebastian Schuchmann, “History of the first AI Winter - Towards Data Science,” 2019.
[4] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,” Nature, vol. 323, no. 6088, pp. 533–536, 1986, doi: 10.1038/323533a0.
[5] “Alvey: the betrayal of a research programme | New Scientist.”
[6] “House of Lords - AI in the UK: ready, willing and able? - Artificial Intelligence Committee.”
[7] B. Butler, “In Cloud Computing Price War Amazon and Google Drop Rates Again | CIO,” 2013.
[8] Jackson Joab, “Price war! Amazon cuts cloud costs to counter Google | Computerworld,” 2014.
[9] S. Ranger, “Microsoft, AWS and Google may have just started the next cloud computing price war | ZDNet,” 2017.
[10] A. N. Bloomberg, “Artificial Intelligence: Reality vs Hype - YouTube,” 2019. [Online]. Available: [Accessed: 23-Jan-2020].
[11] W. Knight, “Reinforcement learning,” Technology Review, vol. 120, no. 2. Massachusetts Institute of Technology, pp. 32–35, 01-Mar-2017, doi: 10.4249/scholarpedia.1448.
[12] F. Woergoetter and B. Porr, “Reinforcement learning,” Scholarpedia, vol. 3, no. 3, p. 1448, 2008, doi: 10.4249/scholarpedia.1448.
[13] Jia Deng, Wei Dong, R. Socher, Li-Jia Li, Kai Li, and Li Fei-Fei, “ImageNet: A large-scale hierarchical image database,” 2009, doi: 10.1109/cvprw.2009.5206848.
[14] “DARPA Grand Challenge - Wikipedia.”
[15] “Tesla Model 3 Driver Claims Autopilot Saved Lives In This Accident.”
[16] “Watch Tesla Model 3 On Autopilot Avoid Crash With Distracted Driver.”
[17] “Watch Tesla Model 3 Autopilot Swerve To Avoid Car Passing On Shoulder.”
[18] S. O’Kane, “Tesla hit with another lawsuit over a fatal Autopilot crash - The Verge,” 2019.
[19] J. Barnett, “The Army working on a battlefield AI ‘teammate’ for soldiers,” 2020.
[20] “Catholic Church joins IBM, Microsoft in ethical A.I. push | Fortune.”
[21] “Algorithmic Accountability Act of 2019 Bill Text.”
[22] J. Emspak, “World’s Most Powerful Particle Collider Taps AI to Expose Hack Attacks - Scientific American,” 2017.
[23] J. McCormick, “Health Systems Look to AI to Prevent Sepsis Deaths - WSJ,” 2020.
[24] Jared Council, “Bayer Looks to Emerging Technique to Overcome AI Data Challenges - WSJ,” 2020.
[25] W. Hählen and S. Kapreillian, “RPA for Tax | Deloitte Global,” Deloitte Tax Solutions.

Did you find this useful?