Definition
Model weights are the trainable parameters of a machine learning model that determine how input features are transformed into predictions.
- They represent the strength or importance of each input feature in the model.
- During training, weights are adjusted (optimized) to minimize loss and improve predictions.
Examples by Model Type
- Linear Regression
$y = w_1x_1 + w_2x_2 + b$
- $w_1, w_2$ = model weights.
- They tell how much each feature $x_1, x_2$ contributes to prediction $y$.
- Example: If $w_1 = 200$, every extra square foot adds $200 to house price.
- Neural Networks
- Each connection between neurons has a weight.
- Forward pass = multiply inputs by weights, apply activation.
- Training = backpropagation adjusts weights to reduce error.
- Example: Image classification → weights in CNN filters learn patterns (edges, textures).
- Logistic Regression
- Weights define the log-odds contribution of each feature.
- Positive weight → feature increases probability of positive class.
- Negative weight → feature decreases probability.
How Weights Are Learned
- Start: initialized randomly or with heuristics.
- Training loop:
- Forward pass → compute prediction.
- Loss function → compare with true label.
- Backpropagation → compute gradients of loss w.r.t. weights.
- Optimizer (SGD, Adam) → update weights.
Why Weights Matter
- They store the knowledge the model has learned.
- Saving/loading a model = saving/loading its weights.
- In transfer learning, we often freeze weights in the encoder and only fine-tune the last layers.
Example
- Spam email classifier:
- Weight for keyword “free” = +2.5 (strongly increases spam probability).
- Weight for “invoice” = -1.0 (decreases spam probability).
Summary
Model weights = the parameters that a model learns during training.
They control how features influence predictions, are updated via optimization, and ultimately capture the model’s learned knowledge.
