Loss Functions

Loss functions measure how well a neural network’s predictions match the expected outputs. They guide the optimisation process by providing a scalar value to minimise during training.

The athena library provides several commonly used loss functions for different types of problems.

Available Loss Functions

• Binary Cross Entropy Loss
• Categorical Cross Entropy Loss
• Mean Absolute Error Loss
• Mean Squared Error Loss
• Negative Log Likelihood Loss
• Huber Loss

Overview

Choosing the Right Loss Function

• Binary Classification: Use BCE (Binary Cross Entropy)

• Multi-class Classification: Use CCE or NLL

• Regression: Use MSE, MAE, or Huber

Loss Function Properties

All loss functions in athena:

• Extend base_loss_type

• Implement a compute method

• Work with differentiable array_type for automatic gradient computation

• Support batch processing

Usage Example

use athena__loss
use athena__network

! Create a network
type(network_type) :: net
type(mse_loss_type) :: loss

! Initialise loss function
loss = mse_loss_type()

! Use in training
call net%train(train_data, train_labels, loss=loss)

Creating custom loss functions

The athena library is designed with extensibility in mind, allowing users to create custom loss functions by extending the base_loss_type.

See the tutorial: Creating Custom Loss Functions