When we ask an explainable AI (XAI) tool to explain why a computer vision model classified an image as a "cat," it typically acts like a spotlight. It highlights the whiskers, the ears, and the paws. This approach treats neural network classification as a purely additive process: the more positive evidence a model finds, the higher its confidence climbs.
But human reasoning is rarely that one-sided. If you see an animal with stripes, you might think it is a zebra—until you notice it has a long trunk. The presence of the trunk does not just add evidence for an "elephant" label; it actively rules out the zebra. Similarly, you might identify a landscape as a desert not because of any specific object present, but precisely because of what is missing: the total absence of trees and water.
Neural networks operate under this exact same tug-of-war. A model's final prediction is shaped as much by what is missing or suppressed as by what is present. However, traditional interpretability tools completely miss this dynamic. To truly understand, debug, and trust deep learning models, we must look at the flip side of attribution: negative evidence.
The Two Sides of Negative Evidence
In modern XAI literature, negative evidence generally manifests in two distinct ways: inhibitory features (things that are present but argue against a prediction) and pertinent negatives (things that must remain absent for a prediction to hold true).
Input Feature Dynamics
├── Present Features
│ ├── Positive Evidence (Supports the class, e.g., whiskers for "cat")
│ └── Inhibitory Features (Contradicts the class, e.g., a trunk rules out "zebra")
└── Absent Features
└── Pertinent Negatives (Must be missing to maintain the class, e.g., no trees for "desert")
1. Inhibitory Features
As highlighted by Sin-Han Kang et al. in Interpreting Undesirable Pixels for Image Classification on Black-Box Models (iccv 2019), competing features within an input constantly suppress or interfere with a model's confidence in the correct class. When a model is confused between two visually similar categories, it is often because an inhibitory feature is driving the target probability down.
Early foundations like Layer-Wise Relevance Propagation (LRP), introduced by Alexander Binder et al. in On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation (PLoS 2015), accounted for this by allowing negative relevance values to flow backward through the network. These negative values explicitly flag input regions that actively contradict a classification hypothesis. By identifying and removing these "distractor" features, as demonstrated by Yipei Wang et al. in "Why Not Other Classes?": Towards Class-Contrastive Back-Propagation Explanations (neurips 2022), researchers can significantly boost a model's confidence in the correct prediction.
2. Pertinent Negatives (Reasoning by Absence)
The second facet is the complete omission of a feature. In Explanations based on the Missing: Towards Contrastive Explanations with Pertinent Negatives (neurips 2018), Amit Dhurandhar et al. formalized the concept of Pertinent Negatives (PN). While a "Pertinent Positive" is the minimal set of features that must be present to justify a decision, a Pertinent Negative is the set of features or empty spaces that must remain absent to prevent the model from switching to a different classification.
For instance, a blank space in a handwritten digit "7" is vital; if you fill in that empty space, the model will instantly reclassify it as a "9".
Why Traditional Saliency Maps Fail
If negative evidence is so critical to model logic, why don't standard heatmaps show it?
Lars Nieradzik addresses this fundamental blind spot in Reliable Evaluation of Attribution Maps in CNNs: A Perturbation-Based Approach (arxiv 2024). Traditional attribution methods operate on the assumption that model decisions are strictly driven by the presence of pixels. When a model relies on an absence—such as identifying a landscape as a desert because trees are missing—there are no positive salient pixels to highlight. Standard saliency maps simply return blank or noisy outputs, rendering them ineffective.
Furthermore, popular "perturbation-based" XAI tools (which test feature importance by blurring or removing parts of an image) run into a mathematical wall when dealing with absence. As observed by Derick M. Oliveira et al. in Explaining Black-Box Automated Electrocardiogram Classification to Cardiologists (CinC 2020), you cannot apply noise, masking, or perturbations to a signal component that is already completely missing. If a medical model diagnoses a patient as healthy because a specific pathological wave is absent from their ECG, perturbation tools have nothing to modify, leaving the clinical reasoning unexplained.
Mechanisms for Capturing Negative Evidence
To overcome these limitations, researchers have developed specialized frameworks that explicitly isolate and map negative evidence.
Disentangled Backpropagation
Instead of letting positive and negative signals cancel each other out during the backward pass, some methods explicitly separate them. Adria Ruiz et al. developed Disentangled Masked Backpropagation (DMBP) in Generating Attribution Maps with Disentangled Masked Backpropagation (iccv 2021) to treat neural networks as collections of linear pieces. DMBP mathematically isolates features that support a prediction from those that contradict it, preventing irrelevant noise from muddying the explanation.
On the architecture side, Moritz Böhle et al. introduced B-cos Vision Transformers in Holistically Explainable Vision Transformers (arxiv 2023), altering the network's internal transformations so that inhibitory signals are preserved and faithfully visualized from the first pixel to the final layer.
Counterfactual Boundary Crossing
Another highly intuitive way to explain an absence is to show what it would take to turn that absence into a presence.
[Original Input: Face without Glasses]
│
▼ (ExplainGAN Transformation)
[Transformed Input: Face with Glasses] ───► Crosses Decision Boundary to "Eyeglasses" Class
│
▼
Explanation: "The absence of eyeglasses is why this was classified as a normal face."
In ExplainGAN: Model Explanation via Decision Boundary Crossing Transformations (eccv 2018), Pouya Samangouei et al. utilized a generative framework to visually transform an image until it crosses the model's decision boundary. By generating the exact features needed to change a classification (e.g., adding eyeglasses to a face), it reveals what the model felt was missing.
Similarly, Joseph Halpern et al. formalized this causally in Explaining Image Classifiers (KR 2024), defining an explanation of absence as the minimal set of counterfactual changes required to trigger a positive detection.
The Danger of Unmonitored Absence: Lazy Learning
Understanding negative evidence is not just an academic exercise; it is a critical requirement for model debugging due to a phenomenon known as lazy learning.
When a model is trained on a binary or multi-class dataset, it often takes the path of least resistance. As documented by Antonio Longa et al. in Explaining the Explainers in Graph Neural Networks: a Comparative Study (ACM Comput. Surv. 2025), a model might learn to recognize a specific motif for Class A, and then assign Class B to any input that simply lacks that motif. It never actually learns what Class B looks like; it only learns that Class B is "not Class A."
This creates a massive vulnerability. Sayan Saha et al. warn in Certified Evaluation of Model-Level Explanations for Graph Neural Networks (iclr 2026) that when a model relies heavily on negative evidence reasoning, traditional evaluation metrics that look for positive feature highlights become completely misleading. The model outputs a high classification score, but the explanation motifs do not represent any real pattern the model used for positive identification.
Without tools designed to capture negative feature signaling, this lazy learning remains entirely invisible to developers, leaving models fragile to real-world deployment where an entirely unrelated third class might also lack the features of Class A.
Key Foundational Work
Defining Salience in Case-Based Arguments (1989) — contributed 2 concept elements to this direction.