VarGrad

Bases: MultitaskExplainerMixin, AttributionExplainer

VarGrad is a gradient-based attribution method that computes the variance of input gradients under random perturbations. Unlike methods that average gradients (e.g., SmoothGrad), VarGrad focuses on capturing the sensitivity and variability of the model's response to local perturbations in input space.

The resulting attributions reveal regions where the gradient signal is consistently volatile, thus potentially highlighting areas where explanations may be less reliable or more fragile.

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.
• Compute the element-wise variance of the gradient values across these samples.
• Aggregate the result per token (e.g., by norm with the input) to get the final attribution scores

Reference: Richter et al. (2020). VarGrad: A Low-Variance Gradient Estimator for Variational Inference. Paper

Examples:

>>> from interpreto import VarGrad
>>> method = VarGrad(model, tokenizer, batch_size=4,
>>>                     n_perturbations=50, noise_std=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
granularity	Granularity	The level of granularity for the explanation. Options are: ALL_TOKENS, TOKEN, WORD, or SENTENCE. Defaults to Granularity.WORD. To obtain it, from interpreto import Granularity then Granularity.WORD.	WORD
granularity_aggregation_strategy	GranularityAggregationStrategy	how to aggregate token-level attributions into granularity scores. Options are: MEAN, MAX, MIN, SUM, and SIGNED_MAX. Ignored for granularity set to ALL_TOKENS or TOKEN.	MEAN
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
input_x_gradient	bool	If True, multiplies the input embeddings with their gradients before aggregation. Defaults to True.	True
n_perturbations	int	the number of interpolations to generate	10
noise_std	float	standard deviation of the Gaussian noise to add to the inputs	0.1