Neural Network
A neural network is the mathematical weighting system at the core of modern AI. It adjusts thousands of numerical parameters through a process called backpropagation until the system can reliably produce desired outputs from given inputs — for example, matching faces to names. The term 'neural network' is a marketing metaphor: it suggests a brain, but it is a chain of matrix multiplications. Professor Jiang's point: the fancy name is designed to make a simple optimization process sound like consciousness.
The structure of a neural network is layers of nodes connected by weighted numerical values. Input data flows through these layers; at each node, the weighted sum of incoming signals passes through a mathematical activation function and moves to the next layer. The final layer produces an output — a classification, a prediction, a text token. Training adjusts all the weights to minimize the difference between the network's outputs and the correct answers in the training data.
The key process is backpropagation: after each incorrect output, the error is propagated backward through the network, and each weight is adjusted slightly in the direction that reduces the error. Repeat millions of times across millions of examples, and the weights converge on values that reliably produce correct outputs. This is the entirety of "learning" in machine learning — it is gradient descent on a loss function, not cognition.
The naming conventions are deliberately evocative. "Neural network" suggests neurons and brains. "Deep learning" (referring to networks with many layers) suggests profundity. "Weights" and "parameters" sound like judgment and experience. "Training" sounds like education. None of these metaphors accurately describe the underlying mathematics. Professor Jiang argues this linguistic strategy is not accidental: it is an occult move — using names to summon a reality that does not yet exist.
Neural networks have a critical vulnerability: they are black boxes. Once trained, the internal weight values are not interpretable by humans. We can observe inputs and outputs but cannot read what the network learned. This is why neural networks can be brittle — a slight change in input that would be trivial to a human can catastrophically fail the network, because the network learned a statistical pattern, not a principle.
Related articles
Related terms
Frequently asked questions
What is a neural network? ＋
A neural network is a multi-layered system of weighted numerical values that learns to produce correct outputs from training data by adjusting its weights through backpropagation — a process of back-calculating errors and nudging each weight in the right direction. It is called 'neural' and described as 'learning' for marketing reasons; the underlying process is mathematical optimization, not cognition. It has no understanding of what its inputs mean.
Why is a neural network called a black box? ＋
Once trained, the internal weight values of a neural network are not interpretable by humans. We can observe what goes in and what comes out, but we cannot read what rule or pattern the network learned. This makes neural networks unreliable in edge cases: they may have learned a spurious correlation rather than a genuine principle. A self-driving car's neural network might only recognize pedestrians in crosswalks — and fail fatally when someone jaywalks.