Compute mutual information (epistemic uncertainty).
I[y;θ|x] = H[y|x] - E_θ[H[y|x,θ]]
This measures the information gained about the prediction by
observing the model parameters.
- Parameters:
predictions (Tensor) – Tensor of shape (num_samples, batch_size, num_classes)
- Return type:
Tensor
- Returns:
Mutual information of shape (batch_size,)