Saturday, April 08, 2017

What is Artificial Intelligence?


bruegel |  The specific term “artificial intelligence” was first used by John McCarthy in the summer of 1956, when he held the first academic conference on the subject in Dartmouth. However, the traditional approach to AI was not really about independent machine learning. Instead the aim was to specify rules of logical reasoning and real world conditions which machines could be programmed to follow and react to. This approach was time-consuming for programmers and its effectiveness relied heavily on the clarity of rules and definitions.

For example, applying this rule-and-content approach to machine language translation would require the programmer to proactively equip the machine with all grammatical rules, vocabulary and idioms of the source and target languages. Only then could one feed the machine a sentence to be translated. As words cannot be reduced only to their dictionary definition and there are many exceptions to grammar rules, this approach would be inefficient and ultimately offer poor results, at least if we compare the outcome with a professional, human translator.

Modern AI has deviated from this approach by adopting the notion of machine learning. This shift follows in principle Turing’s recommendation to teach a machine to perform specific tasks as if it were a child. By building a machine with sufficient computational resources, offering training examples from real world data and by designing specific algorithms and tools that define a learning process, rather than specific data manipulations, machines can improve their own performance through learning by doing, inferring patterns, and hypothesis checking.

Thus it is no longer necessary to programme in advance long and complicated rules for a machine’s specific operations. Instead programmers can equip them with flexible mechanisms that facilitate machines’ adaptation to their task environment. At the core of this learning process are artificial neural networks, inspired by the networks of neurons in the human brain. The article by The Economist provides a nice illustration of how a simple artificial neuron network works: It is organized in layers. Data is introduced to the network through an input layer. Then come the hidden layers in which information is processed and finally an output layer where results are released. Each neuron within the network is connected to many others, as both inputs and outputs, but the connections are not equal. They are weighted such that a neuron’s different outward connections fire at different levels of input activation. A network with many hidden layers can combine, sort or divide signals by applying different weights to them and passing the result to the next layer. The number of hidden layers is indicative of the ability of the network to detect increasingly subtle features of the input data. The training of the network takes place through the adjustment of neurons’ connection weights, so that the network gives the desired response when presented with particular inputs.

The goal of the neural network is to solve problems in the same way that a hypothesised human brain would, albeit without any “conscious” codified awareness of the rules and patterns that have been inferred from the data. Modern neural network projects typically work with a few thousand to a few million neural units and millions of connections, which are still several orders of magnitude less complex than the human brain and closer to the computing power of a worm (see the Intel AI Documentation for further details). While networks with more hidden layers are expected to be more powerful, training deep networks can be rather challenging, owing to the difference in speed at which every hidden layer learns.

By categorising the ways this artificial neuron structure can interact with the source data and stimuli, we can identify three different types of machine learning:
  • Supervised learning: the neural network is provided with examples of inputs and corresponding desired outputs. It then “learns” how to accurately map inputs to outputs by adjusting the weights and activation thresholds of its neural connections. This is the most widely used technique. A typical use would be training email servers to choose which emails should automatically go to the spam folder. Another task that can be learnt in this way is finding the most appropriate results for a query typed in a search engine.
  • Unsupervised learning: the neural network is provided with example inputs and then it is left to recognise features, patterns and structure in these inputs without any specific guidance. This type of learning can be used 
to
 cluster
 the 
input
 data
 into
 classes
 on
 the
 basis
 of 
their
 statistical 
properties
 It is particularly useful for finding things that you do not know the form of, such as as-yet-unrecognised patterns in a large dataset.
  • Reinforcement learning: the neural network interacts with an environment in which it must perform a specific task, and receives feedback on its performance in the form of a reward or a punishment. This type of learning corresponds, for example, to the training of a network to play computer games and achieve high scores.
Since artificial neural networks are based on a posited structure and function of the human brain, a natural question to ask is whether machines can outperform human beings. Indeed, there are several examples of games and competitions in which machines can now beat humans. By now, machines have topped the best humans at most games traditionally held up as measures of human intellect, including chess (recall for example the 1997 game between IBM’s Deep Blue and the champion Garry Kasparov), Scrabble, Othello, and Jeopardy!. Even in more complex games, machines seem to be quickly improving their performance through their learning process. In March 2016, the AlphaGo computer program from the AI startup DeapMind (which was bought by Google in 2014) beat Lee Sedol at a five-game match of Go – the oldest board game, invented in China more than 2,500 years ago. This was the first time a computer Go program has beaten a 9-dan professional without handicaps.