Neural Network
A neural network is a computational model built from layers of simple units connected by numeric weights. Nothing in it is programmed by hand: the behaviour comes from tuning those weights on examples.
How it works
Input passes through layers. In each layer numbers are multiplied by weights, summed and pushed through a simple non-linear function. No single unit "understands" anything; the behaviour emerges from the combination.
Training is weight-fitting. The model is shown an example, its answer is compared with the correct one, the error is computed and every weight is nudged to reduce it. Repeat that millions of times and you have a working network.
Hence the key consequence: explaining why a network gave a particular answer is usually impossible. The knowledge sits nowhere specific; it is spread across billions of numbers.
What to keep in mind
- Data matters more than architectureThe quality and composition of the training set drive the result more than the network's design.
- Bias is inheritedWhatever appeared systematically in the data reappears in the answers, including what nobody wanted.
- Training and running differTraining is expensive and happens once. Running a finished model costs orders of magnitude less.