Model architecture is the chosen form of a parameterized function used to approximate an unknown transformation from inputs to outputs. It fixes what can be adjusted during training, but not the eventual values of those adjustable parameters. In the simplest case, the architecture is a weighted sum of inputs plus a bias. section 1.1 section 1.2
Architecture selection comes before training. First choose a function suited to the problem; then estimate its weights from input-output training pairs. After training determines those weights, the completed function can process new inputs during inferencing. Success depends both on an appropriate architecture and on training-data quality and quantity. section 1.2 section 1.3.5
A threat-score model
For inputs such as hardness and sharpness, a three-parameter architecture computes a threat score using two input weights and a constant bias. The weights scale the inputs; the bias is not multiplied by either input. Training later adjusts all three parameters so predicted threat scores approach known target values. section 1.3.4 section 1.3.5
| y(hardness, sharpness) = w0 × hardness + w1 × sharpness + b |
This weighted-sum form makes the separation clear: the equation specifies the architecture, while training supplies numerical values for its weights and bias. section 1.1
In neural networks, architecture specifically includes the number of neurons and the connections among them. Along with connection weights, these choices determine the network’s ability to represent functions: increasing neurons and connections increases expressive power. Architecture is usually selected for the nature of the problem, often by reusing a design known to work for similar problems. section 7.1
