Learning – the acquisition of knowledge or skills through study, experience, or being taught.Oxford Dictionary
Learning is one of the more abstract parts of life. The arena of education has developed rapidly over the past few centuries, and is now accessible to many more people.
The theory behind how learning occurs is nowadays properly in the fields of neuroscience and psychology. However, before the advent of modern science, learning was studied primarily with the tools of philosophy.
The earliest example is Plato’s slave-boy analogy. The set-up for this analogy is that a slave-boy is set a geometric puzzle by Socrates. The details of the geometric problem are interesting but irrelevant to the main point of his argument. Socrates merely asks questions, and the slave-boy, by answering, figures out the solution to the problem. It appears that the boy has not been taught, but rather has ‘remembered’. This argument was one of the supports for the long-standing philosophical school of innatism, believing that humans are born with innate knowledge.
In opposition to this are the empiricists, who think that knowledge is gained from experience. This is the theory of ‘tabula rasa’ (blank slate).
Before taking a closer look at the application of these theories to the field of AI, it will be helpful to take a look at one of the more fruitful areas for learning in AI.
In Nate Silver’s award-winning study of prediction, The Signal and the Noise, he gives the example of this chess position, between Garry Kasparov (white) and Deep Blue, a computer (black). This was part of the match that proved for the first time that computers were capable of beating humans at chess, at the highest level. Here Deep Blue is up a rook for a bishop and a pawn, which is usually an advantage in the long run.
However, a human player might look at the board from a slightly different perspective. Here is how a human might ‘see’ the position:
A human chess player might look at this position and observe the two connected passed pawns close to Black’s king and White’s bishop and queen on open diagonals leading to Black’s king. Although a computer might be able to account for positional weaknesses such as these, a human can spot them more intuitively. Let’s give some context for those unfamiliar with chess.
A passed pawn is a pawn with no enemy pawns obstructing its path. In this position White has two passed pawns on the f- and g-files, and they are said to be connected. The reason they are more valuable than other pawns is because once they reach the enemy back rank (the end of the board) they can be promoted to queens, making them extremely powerful, almost always enough to win the game. That’s why the position is favourable for White. Indeed, the modern chess engine Stockfish gives the position a rating of +0.4 for White, and the analysis ends with a win for White.
What I have just summarised for you is a human analysis of the position, containing heuristics.
Heuristics are simple mental shortcuts we use to make decisions.conceptually.org
They can help in many situations, ranging from playing board games to analysing social cues at a party. In chess, humans may not be able to calculate as many moves in advance, but they have the experience of others, stored as subconscious instinct about how certain patterns emerge from underlying complexity. Take an example like stock trading. There are people out there, making money from their talent at picking good stocks. The top traders synthesise information from both their computers and the more qualitative evidence that they have of the market, in order to form a more complete judgment.
The classical philosophy behind learning is fundamentally based around a very human point of view. It is too rigid to account for how computers learn, and struggles to make sense of an inanimate object such as a computer containing knowledge.
Others may rightly point to books and videos as examples of inanimate objects passing on information, but at all times such things serve only as a medium for other people to express themselves.
The way that computers learn is of course not from humans, or even other computers. Humans design the broad scale of structure that a computer has to work with, and the end-point of its efforts, but specifically do not try to tell the computer its strategy. After all, the whole point of AI is to obtain solutions that humans can’t create. With a few exceptions, all computers learn by themselves and from themselves, or from test data. The other very important distinction to draw is that computers generally learn new ways of solving problems, instead of new information about the world. This facet of computer science goes back to the very heart of the discipline, with the search for an algorithm that could solve any mathematical problem.
Most modern learning machines work by using neural networks (seen above), which are loosely analogous to how brains work, or by evolutionary algorithms, which practise artificial selection to allow the best designs to survive. In all cases, the learning is completely bottom-up, an emergent property of the system rather than a series of guidelines and rules built in by the developer.
Such computers appear to display intuition, similarly to how a world-class poker player “just knows” when to go all in. But this intuition is a complete illusion, built on the back of superior calculation. Humans have visualisation, seeing the geometry of situations through the lens of our three-dimensional world. Computers just have their lists of numbers, and yet still consistently do better.
The finished product, an expert chess engine or a virtual home assistant, is essentially a black box. We don’t know what goes on in there. There’s an input system, and an output system. Those are the things we have some degree of control over. However, between those two things, it’s ultimately just connections, tweaked and adjusted until they produce the perfect responses.
- Learning – Lexico.com https://www.lexico.com/en/definition/learning
- Plato – sparknotes.com https://www.sparknotes.com/philosophy/plato/section2/
- The Signal and the Noise (2012) 978-0-141-97565-8
- Heuristics – conceptually.org https://conceptually.org/concepts/heuristics
- Programming a Computer for Playing Chess (1950) https://vision.unipv.it/IA1/aa2009-2010/ProgrammingaComputerforPlayingChess.pdf
- How does AlphaZero play chess? (2017) https://www.chess.com/article/view/how-does-alphazero-play-chess
- How Machines Learn (2017) https://www.youtube.com/watch?v=R9OHn5ZF4Uo&t=109s