Most explainability tools are designed to answer a single question: "What features in this input support the model's prediction?" When a model classifies an image as a "dog," a standard saliency map highlights the fur and ears. However, this provides only half of the story. A complete understanding of a model's logic requires knowing not just what features were present to support a choice, but also what features were missing, or which existing features actively argued against that choice.
Researchers are increasingly focusing on negative evidence—the inhibitory signals and contradictory features that decrease a model's confidence in a specific classification. By identifying these signals, we can diagnose why a model is confused between similar classes, find "inhibitory circuits" that suppress incorrect outputs, and build more faithful evaluations of AI reasoning.
The Saliency Gap
Traditional visualization techniques often "zero out" negative attributions, focusing exclusively on positive evidence. As Michail Mamalakis and colleagues argue in Solving the enigma: Deriving optimal explanations of deep networks, this creates a biased view of model reasoning. If a model is 60% confident an image is a "cat" and 40% confident it is a "dog," a positive-only explanation might show the cat-like ears. But it fails to show the snout or tail that might have been providing evidence against the cat label, effectively confusing the user about why the confidence wasn't 100%.
To bridge this gap, Luisa M Zintgraf and Taco Cohen proposed Prediction Difference Analysis (PDA). By strategically removing patches of an image and observing the change in the model’s output, PDA can generate saliency maps that highlight features providing evidence both for and against a classification. This allows practitioners to see, for example, which parts of an object resemble a different class and are thus "confusing" the model.
Propagating Negative Relevance
One of the most influential frameworks for capturing these bidirectional signals is Layer-wise Relevance Propagation (LRP), introduced by Alexander Binder, Sebastian Bach, and Wojciech Samek. Unlike gradient-based methods that can be noisy, LRP redistributes the prediction score backward through the network layers.
Crucially, LRP allows for negative relevance values. In On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation, the authors demonstrate that these negative values identify "inhibitory features"—input regions that actively contradict a classification hypothesis. For instance, in a medical imaging task, LRP can distinguish between patches that support a diagnosis and those that provide evidence against it, a distinction that Julius Hense and Thomas Schnake leverage in their xMIL framework for histopathology.
Contrastive Explanations by Design
While post-hoc methods like LRP try to extract negative evidence from existing models, some researchers argue that explainability should be built into the architecture itself. Liangzhi Li and Bowen Wang introduced SCOUTER, a classifier that uses a slot-attention mechanism to provide "contrastive" explanations.
SCOUTER doesn't just tell you why an image is a "7"; it can explicitly show you why it isn't a "1." By focusing on the absence of certain features or the presence of contradictory ones, SCOUTER provides a more human-like reasoning process. As noted in Explainable Image Recognition via Enhanced Slot-attention Based Classifier, this is essential for high-stakes domains where distinguishing between highly similar categories (like two different types of tissue in a biopsy) is a matter of identifying what is missing as much as what is present.
Discovering Inhibitory Circuits
In the field of mechanistic interpretability, the search for negative evidence has moved from pixels to neurons. Researchers like Aliyah R. Hsu and Georgia Zhou have developed Contextual Decomposition for Transformers (CD-T) to identify the internal "circuits" responsible for model behavior.
A key finding in Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition is the existence of inhibitory components. These are specific attention heads or neurons that function by suppressing certain outputs rather than promoting them. Traditional discovery tools often miss these components because they only look for positive contributors. However, capturing these suppressive signals is essential for a complete understanding of how a model performs complex reasoning and fine-tunes its confidence.
Debugging with the Attribution Confusion Matrix
If an explanation method cannot represent negative evidence, how can we trust its evaluation of a model's errors? Anna Arias-Duart and Dario Garcia-Gasulla addressed this by proposing the Attribution Confusion Matrix.
By treating feature attribution as a classification task (where a feature is either "relevant" or "not relevant" to a class), they show that methods restricted to positive evidence cannot calculate true negatives or false negatives. Identifying negative feature attributions allows practitioners to pinpoint specific regions that actively distort localization or decrease confidence, as explored in the BSED framework by Michihiro Kuroki and Toshihiko Yamasaki.
Ultimately, the shift toward negative evidence represents a move toward "discriminative" interpretability. As Chris Hamblin explores through Maximally Tense Images, understanding what a neuron suppresses is just as important as understanding what it excites. By visualizing the tension between supporting and contradictory signals, we gain a far more robust picture of how deep neural networks actually decide.