Weight
In Artificial Intelligence, especially in artificial neural networks, a weight is a numerical parameter that determines the strength or importance of a connection between two neurons.
When a neuron receives inputs, each input is multiplied by its weight. The network then sums these weighted inputs and passes the result through an activation function. This mechanism allows the network to prioritize certain features over others.
Simple example
In an image recognition system, a weight may indicate how relevant a visual feature (e.g., an edge, a color, or a texture) is for identifying an object.
- A large weight = feature is very important.
- A small weight (close to zero) = feature has little to no influence.
- A negative weight = feature contributes in the opposite direction.
Role in learning
Training a neural network means adjusting weights to reduce the prediction error. This is done through backpropagation and optimization techniques such as gradient descent, which iteratively fine-tune the weights.
Weights are what make a neural network more than just a series of mathematical operations — they are the memory of the model. Each weight stores how much importance should be given to a particular input signal, and together they encode the knowledge learned during training. Without weights, a neural network would remain a static structure with no ability to adapt to data.
One important property of weights is their dynamic adjustment. During backpropagation, the optimizer tweaks weights slightly after each batch of training data, nudging the model closer to better predictions. Over thousands or millions of iterations, these adjustments accumulate into a finely tuned set of parameters that can generalize to unseen inputs.
In practice, different architectures leverage weights differently. CNNs rely on shared filter weights to detect visual patterns across an entire image. Recurrent networks use weights to maintain memory across time steps. Transformers distribute attention weights across tokens, effectively deciding “which words matter most” when processing a sentence.
There is also a research focus on pruning and compressing weights. Large networks may contain millions of redundant parameters, and techniques like pruning, quantization, or weight sharing aim to reduce memory usage without sacrificing much accuracy. This makes weights not only a computational resource but also a key lever for efficiency in deploying AI models on real devices.
👉 In short: Weights are the core knowledge of a neural network — they encode what the model has learned from the data.
