Machine learning: Surpassing humans in quality? Or still subject to "human error"?


Machine learning is the science of getting computers to act without being explicitly programmed.

Artificial intelligence has been the topic of research for decades, and many businesses have been successfully applying the methods of machine learning for many years to solve complex tasks, which are impossible to tackle with traditional algorithmic solutions. Typically these are problems that are easier for a human than for a computer, requiring human-like knowledge, like recognising objects in a picture, driving a car or suggesting a movie a person might like. Applying machine learning used to require a lot of specialist knowledge and custom development, but it is now becoming commoditised and accessible to everyone.

The basic principle of machine learning is building a model of a problem based on a set of examples, often quite large, in a process referred to as learning or training. Machine learning models broadly fall into two categories: classification models, which sort the inputs into predefined categories (e.g. recognise an object in a photo is a dog) and regression models, which learn a relationship between different attributes of a system (e.g. given product information and  your previous purchases, how likely are you to buy the product). Recently, deep learning methods, inspired by the structure of biological neural networks, have become popular and quite successful at solving problems previously considered too hard for a computer to become better at than a human. A perfect example is the success of AlphaGo, a program playing the board game Go, which beat one of the world’s best players in 2016 and was updated in 2017 to learn by playing against itself, rather than from recorded human plays, surpassing the original in three days.

Machine learning models learn from examples and require training data to provide good results. The underlying mathematical methods work with abstract numbers and the quality of the predictions largely depends on the way real-world inputs are encoded before they are given to the model. Results returned by a model will never be perfect - in a sense, machine learning makes the same kind of mistakes that humans would make, but often surpasses humans in quality by several orders of magnitude. The underlying structure and mechanics of the models also have nothing to do with the problems they are applied to, which makes it quite difficult to understand the reason for a particular result and fix specific issues.

Machine learning is now accessible as a service from multiple cloud services providers, both as generic models applicable to bespoke problems and as services tailored to computer vision, natural language processing and other common applications. They provide hardware acceleration and tight integration with their large data stores to support machine learning applications at scale. All of this makes machine learning easier to use than ever, although some specialist knowledge is likely to still be required in order to achieve good results. 

Sign up to Badger News