Smoothgrad

Bases: MultitaskExplainerMixin, AttributionExplainer

SmoothGrad is an enhanced version of gradient-based interpretability methods, such as saliency maps. It reduces the noise and visual instability often seen in raw gradient attributions by averaging gradients over multiple noisy versions of the input. The result is a smoothed importance score for each token.

Procedure: - Generate multiple perturbed versions of the input by adding noise (Gaussian) to the input embeddings. - For each noisy input, compute the gradient of the output with respect to the embeddings. - Average the gradients across all samples. - Aggregate the result per token (e.g., by norm with the input) to get the final attribution scores.

Reference: Smilkov et al. (2017). SmoothGrad: removing noise by adding noise. Paper

Examples:

>>> from interpreto import Smoothgrad
>>> method = Smoothgrad(model, tokenizer, batch_size=4,
>>>                     n_interpolations=50, noise_level=0.01)
>>> explanations = method.explain(text)

Parameters:

Name	Type	Description	Default
model	PreTrainedModel	model to explain	required
tokenizer	PreTrainedTokenizer	Hugging Face tokenizer associated with the model	required
batch_size	int	batch size for the attribution method	4
device	device	device on which the attribution method will be run	None
inference_mode	Callable[[Tensor], Tensor]	The mode used for inference. It can be either one of LOGITS, SOFTMAX, or LOG_SOFTMAX. Use InferenceModes to choose the appropriate mode.	LOGITS
n_interpolations	int	the number of interpolations to generate	10
noise_level	float	standard deviation of the Gaussian noise to add to the inputs	0.1