keyword
gradient-based feature attribution
Gradient-based feature attribution is an interpretability technique in machine learning that quantifies the importance of individual input features or internal representations to a model prediction by calculating the mathematical gradients of the output with respect to those features. By leveraging backpropagation to measure how sensitive the final prediction is to small changes in each input component, such as a word token or an image pixel, this approach assigns attribution scores that indicate which elements most strongly drove the decision. Because standard gradients can suffer from noise or saturation, advanced variants such as integrated gradients, smooth gradients, and input-times-gradient formulations aggregate gradient information across multiple points or baselines. These methods are widely applied to differentiable architectures, including deep neural networks and transformer language models, to evaluate feature relevance, debug decision processes, and trace how internal information flows from input components to final outputs.
1 item
