through our lens
“The question of whether a computer can think is no more interesting than the question of whether a submarine can swim.” — Dijkstra
Artificial Intelligence, often abbreviated as a.i., is the field of creating machines to perform intelligent tasks. For us a.i. is an evolving set of tools that allow us to solve problems. The question of whether the solution is actually intelligent or not, we think, is irrelevant.
Machine learning is a subfield of a.i. that gives computer systems the ability to learn without being explicitly programmed to perform a task. Instead, a machine learns by examining data and tries to find patterns and generalizations. Most of modern a.i. falls under this subfield.
A neural network is a rudimentary mathematical model of a (human) brain, but by no means a real brain. We like to think of it as a function f(x) = y, a way of mapping input data x to output data y. There are many types of neural networks and as such they can solve various tasks e.g. translating Dutch to English, finding a bird in an image, transforming spoken language into text and much more.
Deep learning is nothing more than a neural network spanning multiple layers (hidden layers), hence the word deep. These layers allow for better abstractions and in turn results in higher quality networks. Both the improved learning algorithms and the enormous amount of data that exists today have resulted in much more powerful networks over the past couple of years. They are indeed a very useful tool in our arsenal.
Whereas neural networks require data to learn, reinforcement learning learns by means of a function or reward signal. An agent or robot explores its environment and after an episode or sequence of actions observers the received reward. By combining reinforcement learning with neural networks i.e. let a neural network determine the reward function, a powerful combination can be achieved. This has been applied in e.g. the game of go by deepmind, but is also common in robotics.
These algorithms try to find an optimal set of parameters in a (potentially) enormous search space. They are inspired by evolution through the use of reproduction, mutation, recombination and selection. Like reinforcement learning, they do not require any data, but also rely on a reward signal called a fitness function and simulations.