Error-margin Analysis for Hidden Neuron Activation Labels

Existing neuron labeling methods are incomplete: Current explainability approaches focus primarily on identifying what activates neurons (recall), but neglect to examine what doesn't activate them, creating a significant gap in our understanding of neural network behavior.

Precision matters as much as recall for AI transparency: The paper introduces "error margin" analysis, arguing that understanding both when neurons fire AND when they don't fire is essential for accurately labeling and interpreting neural network decision-making.

Incomplete neuron understanding could mask problematic behavior: By only examining positive activations, researchers may miss cases where neurons respond unexpectedly to certain inputs, potentially hiding biases or failure modes that could affect AI system reliability and safety.

Better labeling leads to better accountability: More precise neuron characterization enables clearer explanations of how neural networks reach decisions, which is crucial for responsible AI deployment in high-stakes domains like healthcare, criminal justice, and finance.

This work addresses a fundamental explainability blind spot: The paper highlights that truly understanding "what a neuron does" requires a two-part analysis—not just identifying its primary function, but also rigorously documenting its limitations and unexpected responses.