learn/basics
Neural networks
A neural network is a model that learns by adjusting the strengths of connections between many simple units, and it underlies almost all of what is now called artificial intelligence.
A neural network is built from units, sometimes called neurons, that each do something very simple: they take in some numbers, add them up with different weights, and pass on a single number. The units are arranged in layers, so the output of one layer becomes the input of the next. The weights are the strengths of the connections, and at the start they are essentially arbitrary. Training means showing the network examples, comparing what it produced with what it should have produced, and nudging every weight a little in the direction that would have made the answer better. Repeat this many times and the weights settle into values that handle the task.
Nobody writes the rules by hand. What the network ends up doing is a consequence of the examples it saw and the shape it was given, not of instructions someone typed. Depth matters: with many layers, early layers tend to pick up simple patterns and later ones combine those into more abstract ones, which is why networks with many layers are described as deep learning. The same basic machinery is specialised by changing how units are wired together, which is where the named families come from, such as the kinds built for images, for sequences, or for data shaped like networks of relationships.
Why it matters for science
Scientists reach for neural networks when a relationship exists in the data but cannot be written down. Much of science has long worked the other way: propose an equation, fit a few constants, check it. That fails when the pattern is tangled, involves many variables at once, or is simply unknown. A neural network sidesteps the need to state the pattern in advance. Given enough examples of inputs paired with outputs, it can approximate the mapping between them, whether that means predicting a protein's shape from its sequence, spotting a flaw in a telescope image, or estimating a material's properties from its structure.
They also stand in for things that are slow or expensive. A simulation that would take a great deal of computer time can sometimes be replaced by a network trained on the simulation's own output, which then produces similar answers far faster. This makes it practical to search through many more candidate molecules, climates or designs than could ever be simulated one by one. The same flexibility explains the reach of the idea across fields: biology, astronomy, chemistry, medicine and climate science all use the same underlying method, differing mainly in what the data look like.
Limits
A neural network learns from the examples it was given and is only reliable on data that resembles them. Faced with something genuinely new, it will still produce a confident-looking answer. It finds associations, not causes, so it can lean on a shortcut in the data that has nothing to do with the real mechanism. Its reasoning is distributed across many weights and hard to inspect, so a wrong answer is difficult to diagnose. It also inherits whatever bias, gap or measurement error sits in its training data, and offers no warning that it has done so.