A model parameter is a learned numerical value that controls how input features contribute to a prediction. Weights express feature importance, while a bias provides an adjustable offset. section 4.2.1 section 3.3.1
Parameters begin with initial values and are tuned from training data. A loss measures prediction error, a cost aggregates it, and an optimization routine adjusts parameters to reduce that error. section 8.2.3 section 3.3.1
A parameter follows a different lifecycle during training and inference:
- Training repeatedly updates parameters using predictions, targets, loss, backpropagation, and an optimizer. section 8.2.3
- Validation loss helps reveal when further updates stop improving generalization and informs choices such as early stopping. section 8.2.3
- Inference uses previously learned parameters without updating them on new data. section 3.3.1
Linear solubility model
For a molecule represented by 11 RDKit descriptors, the model multiplies each feature by its corresponding weight and sums the results with a bias. Learning adjusts these values so predicted solubility better matches measured solubility. section 4.2.1
Neural-network training
An EGFR activity model receives 2,048-bit Morgan fingerprints and predicts pIC50. PyTorch computes the loss, backpropagates gradients, and calls the optimizer to nudge the network's weights and biases. section 8.2
Expand the equation to see how each feature-specific parameter and the bias contributes to a linear prediction. section 4.2.1








































