The terms artificial intelligence (AI) and machine learning (ML) can be encountered in pretty much any description of a contemporary software-based product or service. He we’ll explain the distinction between the two and look at some of the main applications of the latest technology in the world of machine learning. What is ML currently able to achieve and what are scientists’ future aspirations for the technology?
The Difference Between AI and Machine Learning
Artificial intelligence isn’t really something particularly new. It simply refers to a machine doing something in a way we consider is ‘smart’. This is the result of breaking our thought process down into rules and then writing those rules in software code. The machine’s following of these sets of rules is referred to as AI. There is no concrete line in the sand that delineates the point a ‘standard’ computer programme becomes AI. The approximate rule of thumb is that the term AI is used when there is an extensive enough set of interconnected rules for the end result to create the impression of machine ‘intelligence’.
Artificial intelligence is ‘applied’ to a particular task. The computer is provided with a set of rules and a database of information to which it applies those rules. Driverless cars are an example of applied AI. The computer ‘system’ consists of a huge set of data including a detailed map, weather conditions and potential external influences such as other vehicles, pedestrians or unexpected obstructions that might appear on the road, such as a fallen branch. To execute a journey from point A to point B on the map, rule ‘1’ might be “drive to junction”. Rule ‘2’ – “at velocity ‘x’ maintain a distance of ‘y’ from the car in front”. Rule ‘3’ – “in the event of a moving object entering the road at a perpendicular angle that will lead to crossing trajectories, reduce speed”.
Of course, a working driverless car system requires an extremely wide and deep set of interconnected rules that take into account every potential set of circumstances at any given moment. Which is why, when achieved, and it soon will be, driverless cars will represent a watershed moment in AI advancement. Nonetheless, as impressive an example of the potential of AI technology as driverless vehicles will be, it will still represent machines following, without deviation, a set of pre-programmed rules. Other examples of applied AI are algorithms that automate financial markets trading or products or viewing suggested by Amazon or Netflix and based on your previous user behaviour.
Machine learning is often defined as a sub-category of AI. Perhaps a better way of defining ML is as the advancement of AI. It can be considered as general, rather than applied, AI. The difference is that, in theory, ML systems can be applied to any task. This is achieved through a marriage between algorithms and data.
An ML algorithm attempts to capture the rules along which human brains work when we receive external data input such as sensory information or facts, subconsciously, or consciously, analyse that data, and come to conclusions.
Those conclusions then influence or subsequent beliefs and actions. This is, in essence, the process of learning. Computers have a significant advantage over human brains when it comes to the amount of data they can store and easily access, the speed and accuracy with which they can make calculations and their lack of bias. The limitation, for now, comes from how well we are able to create the base algorithm as a reflection of how our own brains work. They may be slower and more fallible than a computer but they also have an extraordinary depth of complexity we don’t yet fully understand and are therefore unable to replicate fully through software rules.
The result of that limitation is it will be some time before we have to genuinely worry about the risk of ‘singularity’ in ML. This refers to the theoretical tipping point of AI into self-awareness. While our own lack of a complete understanding of the minute intricacies of how the human mind works means it can’t be completely programmed, we do know enough to be able to create an approximation. And that’s what is leading to some of the impressive recent achievements in machine learning.
Neural Networks in Machine Learning
The architecture of neural networks in software mimics the way the human brain classifies and categorises information. This works on the basis of probability. For example, if an animal has fur, we classify it as a mammal. A feedback loop informing a machine learning AI if its classification has been correct or not means its decision making process continues to evolve.
Applications of Machine Learning
Image recognition – one example of machine learning many of us will encounter regularly is image recognition. When you take a photo of a landmark with your smartphone and Google Pictures recognises it and prompts its public upload, it has used machine learning to spot the defining elements of that landmark in your photo, connecting the two. Of course, when your phone’s GPS is switched on this also contributes, by significantly narrowing down the options.
Chinese police were recently able to apprehend a wanted individual because machine learning-powered facial recognition software connected to security cameras at a concert he was attending was able to match his features to a police database. Officers arriving at the concert were then assisted in picking him out from the crowd by ‘smart’ sunglasses they were wearing that also used facial recognition software.
Targeted Marketing – ecommerce sites such as Amazon, social media platforms and other digital platforms for marketing use machine learning to better target ads to users. AI will inform an add platform that someone who has bought a football shirt, or ‘Liked’ several football related groups is likely to be interested in other football-related products.
Machine learning adds an extra layer to this initial information. If the user does not respond to other football-related ads, the platform will ‘learn’ that the previous purchase was not a general trend but a relative one-off. This could be because the shirt was a present for someone else or they only buy one football shirt a year, because their favourite team’s new model has just been released or the soon-to-be-replaced model went on sale. The platform will learn from the failure of the other ads served and optimise the consumer’s future ad profile. However, the user may be served other football shirt and related ads at the same time the following year when the same friend or family member’s birthday comes around again or they are ready to buy themselves their annual shirt.
Over time, the machine learning software will build a far more accurate and deep consumer profile than would be the case if simply relying on static data such as previous products purchased, groups joined or pages ‘Liked’.
Genomics – genomics is one of the areas of medicine that machine learning is having a particularly strong influence in. Each individual has more than 20,000 genes and there is a high degree of variation within each of those 20,000+ genes. Specific genes in isolation often do not lead to a health issue or disease expression, which can often be the result of complex combinations of genetic factors as well as external influences a patient has been exposed to.
However, machine learning is being used to gradually piece together the complex combinations of patterns in genes, and data on external influences where available, to build a picture of how individual genomics can impact health. A further layer is then added by incorporating data on how individuals respond to different treatments.
Natural Language Processing – machine learning is finally allowing software developers to get to grips with the challenge of creating machines that can communicate with humans in a ‘natural’ way. As anyone who has ever tried to learn a foreign language will be well aware, standard grammar rules only get you so far, even if you can learn them perfectly. There are so many anomalies, exceptions, nuances and combinations that the only way to learn to speak a language in a ‘natural’ way is by gaining experience of all of these non-standardised details and learning through context.
As such, it is only now that advancements in machine learning technology means that serious strides are being made towards truly naturalised language-based interaction between humans and machines. The same can be said of the improving quality of automated translation software.
There are of course many, many more applications of machine learning and new applications and developments are now happening at a frantic pace. While machine learning is taking AI to new levels with each passing month it also still has many limitations. These demonstrate just how intricate the natural learning process of the human brain really is.
At its current pace of development, machine learning holds tantalising prospects in many fields of application. It will within our lifetimes undoubtedly lead to a huge amount of automation in areas that have so far resisted attempts to minimise human input. Managed correctly, the latest technology in the world of machine learning should benefit society immensely across diverse applications from health to the economy.