What is machine learning?

Saying that computers are extremely powerful is an understatement. They have revolutionised our world, and the change is far from over. However, traditional computing relies on something called programmed learning, where the computer’s function and task must be programmed in directly by a human. You only need to pull out a calculator to see that they are capable of performing certain tasks with a rapidity and focus that cannot be matched by humans. However, they are also extremely limited.   Most computers can only do what a human tells them to do, and if you put in the wrong command they will do the ‘wrong’ task. Not only that, but they are also limited to the list of functions that they are programmed to. Artificial intelligence uses a different method of computing known as ‘machine learning’ that enables it to be much more effective at goal-orientated tasks than traditional computing. This has opened up a new frontier of data management and started a new stage in the Information Age.  

How do normal computers work?

  In 1997, IBM’s computer Deep Blue beat the reigning chess master Gary Kasparov. At the time this was seen as a major coup for computers, and there was lots of buzz that our lives would be immeasurably transformed by super-smart computers. While great leaps have been made in computer performance since then, this obviously has not happened yet. Deep Blue was not the watershed moment that it seemed to be.   True, nothing like that had ever been accomplished before, but fundamentally Deep Blue was a testament to the ability of computers to do some serious number crunching, and systematically follow rules. It worked by ‘brute forcing’ every possible outcome from a chess move and working out the best response by an algorithmic process. Nevertheless, it demonstrated that computers were very effective at specific, highly regulated actions. The rules for chess can also be put down relatively easily, but crucially they had to be programmed in by humans.  

What is machine learning?

  AI operates in a different way from what we as computer users are used to, by using a function called ‘machine learning’. Rather than number crunching, AI uses a neural network that is able to ‘learn’. This is generally done by the computer looking at a large number of examples of something and then identifying principles that link each of the examples together. This is modeled on the human brain. A good way to think about this is to think of the CAPTCHA programs. You might have had to click on ‘all the motorcycles’ in a set of pictures to demonstrate that you are ‘not a robot’. This exploits the fundamental weakness of a traditional computer, which would have to be told whether each specific picture had a motorcycle or not.   Humans are very good at performing tasks like image recognition. Even if we have never seen a picture before, we are able to do a lot of pattern recognition. Data scientists who develop AI are trying to replicate the size-to-power computing ability of the human brain. We manage to be highly efficient by being conservative with how we tackle problems. Take a look at a picture of Deep Blue. The massive computing power came at a cost – it couldn’t fit in the same room as Kasparov.   Machine learning is so powerful and efficient because it works by a computer recognising patterns that underlie a set of data by abstracting out and generating rules based on this abstraction. It can then systematically approach new data in a way that is minimal, rather than having to run through each possible outcome. This was put to use in 2017 when Google’s ‘DeepMind Alpha’ project beat ruling Go champion Lee Sedol. Go is a board game popular in China and Korea that has a possible number of legal moves up to 2 x 10170 , making it far more complex than chess in terms of pure numbers. Each move in Go has 250 options and each game typically last for 150 turns. On the other hand chess has 35 possible opening moves, with an average game lasting for 80 moves. Obviously, this is still a massive number – it even has its own fancy name: ‘Shannon’s number’ – but it is still possible to brute force to a degree. DeepMind didn’t just crunch all the possibilities - it worked out effective strategies to play more successfully, mostly without human guidance.  

Truly intelligent AI

  Artificial intelligence is ‘intelligent’ in the sense that it can choose the rules that it follows to complete a task. Rather than a user having to pre-program an endless list of rules, true AI will be able to select and create its own rules and complete a task that way, in a similar way to a human. Artificial intelligence already exists to a degree – if you’ve come from Google to this page then you’ve been helped out by AI. It has had to understand what you’ve typed in, look for relevant content, and put them in an order that makes sense! This would take a human a long time, but AI can do this almost instantly. Of course this may sound a bit scary, but as usual, the reality is lagging behind the theory.   Computers may be able to beat a human at chess and paint an expensive picture, but that doesn’t mean that humans are somehow defunct. Just as photography did not mean the end of painting there will also be tasks that humans are just better suited to than computers. AI isn’t a panacea; it doesn’t actually understand what you want in any meaningful way, it just is able to really efficiently spit out a response to what it thinks you want. The form of response is goal-oriented, where the computer ‘knows’ the ultimate goal, but not the route to achieving it. In fact it is a bit more complicated than that, because AlphaGo relied on a database of Go games played by humans to compare against, making it a form of ‘semi-supervised’ learning. Whereas true AI would be able to find out how to play by purely unsupervised learning that does not require a human to provide the examples.  

New frontiers for AI

  That’s why there will always be certain tasks that humans can just do better. Humans may not be able to beat computers at tasks that involve doing lots of sums very quickly, but we tend to be good at things that computers are unable to do. The sorts of tasks that we excel at generally have rules that are very difficult to codify, either because we do not know them ourselves, or because they appeal to our adaptable nature. Young children are able to understand sentences, form new ones, and generally don’t make too many mistakes. By the time they are older they can do this even better, and understand words from context, or in a foreign accent, or that are pronounced incorrectly. Computers traditionally have a lot of trouble with this - just think how patchy voice recognition services are, even when they have been personalised to you. However, the technology is improving every day, with exciting possibilities being realised.   Take customer service – we are a long way off from AI shop assistants, for example. However at We Build Bots we think that some customer service jobs are better done by computers. Highly repetitive jobs that require 24 hour cover and demand quicker response times are not done particularly well by humans. We tend to want things like breaks, sleep, holidays, and sick pay. An AI chatbot, on the other hand, will work tirelessly.   One of the major obstacles is making a bot that can understand nuances like dialects and mis-spellings. Google carried out some research on how children search online that revealed some of the problems of using an AI system for people of mixed abilities. When asked to find out ‘what do dolphins eat’ many of them simply typed in ‘wat dolfin eat’. This is recognisably close enough to the original query, so wasn’t too much of a problem for Google. However, adapting government or council projects to make sure that they have sufficient accessibility for all members of the public is a significant challenge.   When it comes to customer service there will always be such challenges. But as the technology improves it also means that more and more people will be able to get the help they need from a system that can understand their needs, and give answers, no matter the time of day, or the problem, or the person asking. Of course, there is much more to AI technology than just machine learning, and there are different approaches to computational learning too. However, the fundamentals of using a computer to identify abstract patterns are consistent throughout Artificial Intelligence development and offer exciting potential routes for using computers to revolutionise our world.