Machine-Learning: To Be Feared Or Embraced?

A recent article in the Guardian highlighted a perceived risk from automated decision making processes. These automated processes are based on a technology known as “machine-learning”, and the article drew attention to the controversial experiment where a machine-learning algorithm was given the task of judging beauty pageants. In the experiment it was found that the winning contestant was invariably a white woman. The article could be described as informative or factually incorrect and alarmist depending on your viewpoint, but it raised an important question about whether there is something to fear from automated decision making processes in general. In order to form an opinion, it is important to understand what machine-learning is and how it works.

Arthur Samuel defined machine-learning as “a means of giving computers the ability to learn without being explicitly programmed”. Paraphrasing, it is a means of giving computers the ability to learn from experience. Applications include prediction of house prices in a particular area or the prediction of the likely demand for ice-cream based on the weather. Other applications include classification of articles on the Internet or the classification of a sample of cells from a medical procedure.

There are essentially two types of machine-learning. The first type is generally used for predicting values, such as the likely sale price of a house in a particular area. It does this by collecting information on previous sale prices, so each additional piece of data makes the next prediction more accurate. This type of machine-learning is called Supervised Learning because it is making predictions based on real and accurate historical data called "Training Data". The second type of machine-learning is used for classification rather than predicting values. Classification is very different to predicting values, and is useful in situations where you might want to be be alerted to a sudden increase in the number of a particular type of illness in a given area. It might point to a new environmental factor that needs further investigation. Similarly, classification is very useful in web searches allowing you to search for information relating to the sleeping habits of gerbils. Clearly, the search engine knows nothing about gerbils or their nocturnal habits. The search engine works by classifying articles based on searches by other people, and ranking the results based on the number of times a particular article or link has been followed by different users. This type of machine-learning is called Unsupervised Learning, because there is no initial data. For most of use who use the Internet, it is an indispensable tool, offering almost instant results to research questions that could have taken weeks or months of research only a decade ago.

Machine-learning offers the ability to learn rapidly from statistics, building from history to fine-tune predictions with some obvious benefits. The ability to detect changes in cells that could, in time, develop into something malignant is a major advance and results in significantly better outcomes for patients. Improved weather prediction means better preparedness which saves lives. The other side of that coin is that it is possible to draw an inference between observed data and a possible outcome. There is a saying in statistics that correlation does not imply causation. Hot weather results in increased ice-cream sales. Hot weather also results in increased accidental drownings. At first glance it might appear that there is a link between ice-cream sales and accidental drownings. To a Human, common sense will usually prevail. A machine, however, might make this kind of connection without the benefit of so-called common sense. As ridiculous as this may sound, a recent report has suggested a potential link between degenerative neurological conditions and a particular brush stroke employed by artists. Just like the ice-cream versus accidental drownings situation, there could be many reasons for this correlation, including training or simply following a particular style. Now imagine if life assurance companies were to use this type of report and the implications for their customers.

So, to the question, is machine-learning something to be feared or is this headline-grabbing click-bait? Probably both. With regard to the apparent bias of machine-learning applied in the selection of a beauty pageant, the machine-learning algorithm has no bias. It offers results based on the training based on what it has learned. If it is operating on training data, and all previous winners have been white, it will behave accordingly. The danger is that machine-learning algorithms could accidentally result in enforcing the status-quo, not because of the algorithm but because of the data it is learning from. Online advertisements for jobs, could continue to feature jobs offering different levels of pay depending on your perceived race, gender or post code, accidentally reinforcing prejudices and creating a new caste system.

The issue is not the machine-learning algorithms, the issue is the data used and the ethical and sensible use of that data.