Definition
Plain language
An interpretability trick that measures how much an input mattered by combining how sensitive the output is to it with how big that input's own value was.
As stated in the literature
An attribution method (Shrikumar et al., 2017) that weights a gradient by the corresponding input activation to cancel some noise relative to raw sensitivity; in GBC it underperformed pure gradient (loudness) attribution as a cross-agent blame signal.
Why it matters: It offers a cleaner way to see which inputs drove a model's decision, helping researchers assign blame or credit when diagnosing behavior.
For example, to explain a decision it checks not just how much a single input could sway the output but also how strong that input actually was, so a barely-present feature gets less credit.
Heard on the show
“That second one is a known interpretability method, gradient-times-input, from Shrikumar and colleagues back in 2017.”Episode 181 — How to Backpropagate Blame Through a Team of Chatbots — And When It Backfires