Machine learning is a branch of artificial intelligence in which computational systems learn from their own experiences in order to make decisions. Rather than hardwiring in a system’s responses to inputs, it is designed to work out for itself how to complete certain tasks. The more example data the system is given the better the results will usually be. These systems are especially good at solving problems that would usually require a human brain such as recognising and classifying images, handwriting recognition, facial recognition, natural language processing (understanding spoken and written language including idioms and sentiment, sometimes language translation), self-drive cars, spam email and fraud detection, and medical diagnosis. Despite its representation in popular literature however, machine learning does not really mimic human thought. It is simply used to augment human intelligence and extend our capabilities, often speeding up complex but repetitive tasks or drawing conclusions from data more quickly and accurately than by other methods. Whilst machine learning research began in the 1950s and both theoretical and progress has been accelerating over the time since, it only became feasible to use it widely for real world applications during the last decade when advances in computer hardware and infrastructure made it possible to run these very complex systems in an acceptable time frame; as well as the increasing availability of large amounts of data to train machine learning algorithms.

Generally, there are three ways to build a learning system: “supervised learning”, “unsupervised learning” and “reinforcement learning”. Supervised learning algorithms take in labelled examples and use them to figure out how to identify the labels from features of the data. For example, you might provide the algorithm with a lot of examples of images of tumours, which a qualified clinician has labelled as malignant or benign. The algorithm will compare all of the photographs and figure out features that differ between the malignant and benign tumour (not always features that a human might consider). This is called training. When it is later given an unknown tumour it compares its features with what it has learned, and makes a decision about whether the photograph is more likely to be showing a malignant or benign tumour. These sorts of algorithms are used wherever there is clearly labelled data available, for example in determining what sorts of customers are most likely to default on a loan based on a past pool of defaulting and non defaulting borrowers; predicting traffic flows; identifying objects or recognising faces in photographs; predicting the price a house will sell for given the geographical area and features of the property, and pattern recognition such as identifying flaws in a structure from photographs.

Unsupervised learning is a more complex case, in which the machine attempts to look for patterns in data, without knowing exactly what it is looking for. It does not have an “answer” for comparison, it simply tries to separate pieces of information into groups that might be useful (this is sometimes called “data mining”). For example you might have a lot of photographs of birds, but not know which birds are closely related and which are not. An unsupervised learning algorithm would detect closer similarities between some birds and others and group them into sets of “similar” birds. However, whilst it could detect which group a new bird should belong to, it would not have any way to tell you the name of the group; and it may not match up with the expectations of an ornithologist. This sort of learning is often applied in targeted marketing or recommendations, whether the algorithm is aware that people who watched Game of Thrones often also watched Breaking Bad, and so if you watch the former it might be likely to recommend to you that you try the latter. It is not aware of what the link might be between the films, or why someone who watched one might want to watch the other, just that it is likely. Similarly if you place certain Amazon items in your shopping basket it might recommend that you look at other products that people with similar baskets went on to buy. It is extremely useful in jobs like file compression, where it might for example figure out what elements of an image are essentially repeated and which parts do not provide much information, and reduce the file size based on that analysis.

Some jobs can be done using either supervised or unsupervised messages, such as fraud detection. You might have a list of fraudulent and valid banking transactions that is labelled according to whether the bank staff believe they are fraud, and train a (supervised) machine learning algorithm based on that information. Or you may simply pass into an unsupervised algorithm a lot of trades, some of which are fraudulent, and see which ones it determines have qualities that are in some way different to normal trades, then investigate them for possible fraud.

Finally reinforcement learning algorithms learn from experience over time. The classic use case for these sorts of algorithms is in creating programs that can play chess or go against a master, and eventually win. They are “semi-supervised” as at each stage the algorithm has some information about how it is performing but often not complete information, and they are often used when the route from the current state of events to the desired state of events is not entirely clear. For example in deciding the optimal configuration of traffic lights, where over time the model can learn which actions it can take that will lead to the best flow of traffic (after initially guessing and making some serious mistakes!) They are very useful in robotics, in figuring out how complex biological and chemical interactions take place (for example in pharmaceuticals) and in system management such as balancing load on servers. In fact Google used their learning algorithms to make an energy saving of 40% in their own computer centres[1], an interesting example of learning computers managing their own environment.

Artificially intelligent machine learning is heavily used in a wide range of industries, including healthcare and medicine, car manufacture, education, marketing and advertising, human resources and recruitment, retail and e-commerce, logistics, buildings management. The power of machine learning algorithms can be limited, however, both due to the fact that they often require a significant amount of compute power and very high-end hardware to run training exercises in a reasonable timeframe, and also because they need huge sets of data to learn from. An algorithm designed to recognise images, for example, requires very large numbers of example photographs in order to train to recognise certain features. There are no methods at present by which you can teach machine learning algorithms using abstract concepts. This means that you cannot simply tell an image processing system that if it sees something with four legs and a flat rectangular top that it is likely to be a table - it may need to see several thousand labelled examples before it will do a good job of identifying the next table. If it is not provided enough learning data the algorithm is likely to generate unexpected results, maybe identifying an assault rifle as a table. Similarly it cannot be expected to understand concepts such as justice, democracy or unfairness, and usually does not have awareness of any context to the data it sees. Machine learning tools are generally not adaptable, they understand only the specific problem they were trained for, succeeding only in an unchanging environment, and they cannot be used to solve problems where there is not much suitable data available to train on, such as diagnosing and treating rare diseases.

Whilst this sort of algorithmic failure due to lack of proper training may be benign, insufficient or biased input data can sometimes have very serious results. A self drive car that does not recognise a pedestrian is a clear danger. Furthermore machine learning algorithms inherit any bias that exists in the data they learn from. Many examples of racist or sexist algorithms have been inadvertently unleashed on the public. In 2018 Amazon was forced to retire a recruitment algorithm that was shown to strongly prefer male profiles, having been trained mainly on information about men[2]. Similarly Google was accused of advertising jobs with higher salaries to men than women[3], and one English firm used an algorithm that was instructed to prefer good quality written English in job applications, leading it to infer that people with foreign-sounding names (who were more likely, but not certain, to use English as a second language) should be summarily rejected[4]. With an even more serious impact, an algorithm used in the USA to predict risk of reoffending (COMPAS) was found to be racist, assigning an artificially higher risk to black over white people and leading to some black people being unfairly denied parole[5]. The data used to train these algorithms contained within them societal bias such as information on existing large male/female pay gaps; or the correlated facts that poverty drives crime, that black people tend to be poorer in the USA, and that the justice system is already unfairly biased against black people. Data must be very carefully selected in order to train algorithms ethically, and this is not a trivial problem.

Artificially intelligent machine learning can of course, like so many tools, be used for deliberately damaging purposes. From ethically ambiguous automated targeting and deployment of super weapons, to sensors collecting private information for use in marketing or law enforcement, to invasive and unpleasant advertising experiences, it will always be possible for someone unscrupulous to create algorithms that work in a way we do not desire them to. However the fear in some quarters that machines will be created with superhuman intelligence and a desire to wipe out human life is highly overblown. The state of the art for creating an algorithm that can truly act as an independent sentient entity is still far away, and it is likely always to be a matter of triviality to write in a clause that says “harm no humans”. Incompetence is by far the greater risk in most cases than evil design.

Another significant problem is that a large number of machine learning tools, especially “deep learning” systems that train on enormous data sets, work in mysterious ways - the so-called “black box”. The artificially intelligent system will ponder a problem and produce an answer, but does not give information on how the decision was made, and the same decision may not always be reached in response to the same input. In some fields a proper audit trail to explain decision making is a legal or ethical requirement. Identifying a table as a rifle may be an embarrassment for image labelling software, but losing significant client money in financial trades or misdiagnosing a patient could be extremely problematic, especially if the logic that caused the event is unknowable.

Machine learning can also provide tools that replace human work, leading to the potential for economic or personal distress to certain groups, usually people whose jobs involve repetitive and easily recreatable workloads. From 2019 to 2022 machines and algorithms in the workplace were predicted by the World Economic Forum in 2018[6] to displace 75 million jobs worldwide, but to create a further 133 million new roles; creating overall 58 million new jobs and therefore decreased unemployment. However alongside that is a shift in the sorts of roles available for people to do, and a huge requirement to retrain employees who work in replaced functions. This is likely to lead to a huge shift in the way jobs are allocated, their permanency, field of industry and locations that may lead to negative social and working consequences for some. Around 50% of companies expect their full term workforce to shrink, and 40% expect it to grow, meaning that some people are likely to be forced into moving jobs or careers. However Artificial Intelligence is expected to provide a boost to the world economy estimated at $15.7 by 2030 [7], so whilst a difficult adjustment it is to be hoped that this revolution will be a large net benefit.

However despite its shortfalls, machine learning is a fascinating, diverse and fast moving field with the potential to alleviate many world problems in science, technology, personal wellbeing and medicine amongst others, save time and money, and even help save the environment. If managed well and ethically it has a high capacity to change the word for the better.


References:

[1] R. Evans and J. Gao, DeepMind, “DeepMind AI Reduces Google Data Centre Cooling Bill by 40%” July 2016

[2] Reuters: “Amazon scraps secret AI recruiting tool that showed bias against women”, Oct. 2018

[3] MIT Technology Review, “Probing the Dark Side of Google’s Ad-Targeting System”, July 2015

[4] Claire Jamie-Lee Nolan, ENSafrica: “Can algorithm bias result in employment discrimination?”, April 2019

[5] J. Dressel and H. Farid, Science Advances Vol. 4, no. 1, “The accuracy, fairness, and limits of predicting recidivism”, Jan 2018

[6] World Economic Forum, “The Future of Jobs 2018”

[7] PwC, “PwC’s Global Artificial Intelligence Study: Exploiting the AI Revolution”