Guest Article by Sohan Seth & Daga Panas

University of Edinburgh, School of Informatics

Why do we need machines to learn?   

You may have noticed that the concept of intelligent machines is cropping up almost everywhere these days. News, social media, popular culture, and even government initiatives seem to be full of references to Artificial Intelligence (AI) or machine learning.

Banded by some as a boon to humanity, by some as a threat, intelligent machines have even been made protagonists in a number of Hollywood movies, usually as a scary villain (Terminator, Matrix, or Ex Machina).

While the doom-and-gloom scenario is certainly exaggerated, it is undeniable that AI is fast becoming an indispensable tool in our daily lives. Smart home systems, self-driving cars, fraud detection and even autocorrect are some of the situations where AI and machine learning are changing the way we live (although autocorrect might have been a constant source of frustration for many).  

We sure don’t see many robots in the classrooms yet, so what do we actually mean by the term machine learning?

Machine learning is a discipline of computer science that deals with designing computer programs on different principles than the ‘traditional’ way of painstakingly laying out steps for the computer to step through, and conditions to check (e.g., if it is black and white and has stripes then it is a zebra).

Much like a human brain, a silicon ‘brain’ can also learn to perform a task simply by seeing enough examples (e.g., images of zebra) of it. All that is required is a program with adjustable knobs such that the program can correctly replicate the given examples (e.g., recognizing zebra in images) by adjusting these knobs. We call this program a model, the knobs are called parameters, the examples the referred to as data, and adjusting the knobs becomes the training process. 

To make it a bit more concrete, let’s say we wanted to write a program recognising zebras in pictures. If we were to do it by explicit instruction, it is hard to imagine where to even begin – after all, we humans do it automatically, with little conscious insight into the details. We might start from the idea that there need to be black and white stripes, and four legs, and a tail… but quite quickly, if you think about it, we run into a multitude of problems. For instance, in some poses we might only see two legs and no tail, so we would have to add special conditions to include these cases; or we might encounter pictures of lemurs who also have stripes and four limbs, so we would have to add more conditions to exclude those cases; and before we know it we are in a  jumble of ‘if-then’ conditions that is impossible to manage. 

This is exactly where machine learning shines. All we need to do is collect a good collection of images, each annotated as to whether there is a zebra in it or not; and show these examples to the model. During the training process, the model will receive these examples, one after another, each time using the parameters to make a decision. Should the calculations lead to a wrong answer, the parameters of the model will be modified to be more likely to output the right answer. 

 This method of designing software, naturally, has both benefits and costs. The obvious benefit, as in the above example, is that it can lead to excellent performance even in the absence of explicit instructions (not to mention we can kick back and relax while the machine is learning). Another benefit is that machines can then perform such skill-demanding tasks at a scale or speed far surpassing human capabilities (e.g., if we want to find all the zebras in a savanna).  

The costs are perhaps less obvious. One of the major concerns is the lack of transparency, the very same notion that prevents us from writing down exactly what steps human brains take to arrive at a decision. The numbers of parameters of a model can easily go into billions, and if at times the resulting program gives counter-intuitive answers (e.g., it is okapi not a zebra), it can be very difficult to trace back why. Another drawback is the ‘data-hunger’ of these algorithms; it took decades of accumulation of digital data for AI to get to where it is now.  

 The bottom line, as with any new, powerful technology, is the old adage:

‘with great power comes great responsibility’.

And it certainly is a very powerful technology. There are so many examples of where AI has helped us and where it promises to provide help that trying to list them all would seriously strain readers’ patience. To highlight the possibilities opened by machine learning, consider for example the task of understanding land cover from satellite images, as might be needed in assessing deforestation or estimating crop yield. Since such an application would require regular surveying of images from many thousands of kilometres, it is impossible for a single human to perform, but – once trained – trivial for a model. 

A similar situation where the amount of data to process is just astronomical is… well, astronomy. If we want to find habitable planets in the universe, there are as many places to look as there are stars in the sky – quite literally! A different yet example of scale and a slightly different advantage to consider is analysing medical X-ray images. In such a situation a machine might be able to outperform an individual human only because it is able to learn faster. Having access to any and all images ever produced, it can reach a level that is hard to achieve even by the most dedicated professional – after all, a model never sleeps. 

Significant efforts are being made by large organizations and machine learning enthusiasts alike to help the general public understand this topic better.

There are accessible machine learning courses and codes (https://developers.google.com/machine-learning/intro-to-ml)

and even bite-sized articles (https://towardsdatascience.com/machine-learning/home)

that can be easily digested over a coffee.

Learn more about what the machines are learning.