A neural network is a machine learning model made of connected computational units whose adjustable parameters are trained to map inputs to outputs.
Training adjusts parameters
A network transforms inputs through connected operations. During training, an optimisation process adjusts parameters, often called weights, to reduce a chosen error measure. During inference, the trained network produces an output for a new input.
Processing more inputs during ordinary use does not necessarily change its weights. The comparison with biological neurons is an inspiration, not a claim that the model works like a human brain.
Example: classify product images
A network trained to distinguish two products might learn the background used in their photographs rather than the product itself. It can score well on similar pictures and fail when the background changes.
Evaluate on separate images that represent the intended lighting, angles and environments. Inspect failures to determine whether the model learned a useful relationship.
Match complexity to the problem
Compare with simpler models and account for training data, computing cost and maintenance. More layers or parameters do not guarantee better results on the task.
Google’s machine learning course introduces neural network components. Deep learning uses networks with multiple representation layers, while machine learning includes many other model families.