Deep learning models are notoriously "black boxes." Feature attribution methods—tools designed to highlight which parts of an input were most important for a model's prediction—aim to open these boxes and provide interpretability. However, the field currently faces two major hurdles. First is the faithfulness gap: many popular methods produce visually pleasing heatmaps that do not actually correlate with the model's underlying decision-making process. Some "interpretable" methods even produce identical results when the model's weights are completely randomized. Second are resolution and architecture constraints: previous state-of-the-art methods, like Information Bottleneck Attribution (IBA), work by placing a "bottleneck" on latent (hidden) layers. To map these deep features back to input pixels, they rely on upsampling and averaging, which results in blurry explanations and typically only works well for Convolutional Neural Networks (CNNs).
To address these limitations, researchers have developed InputIBA, a novel feature attribution method that identifies which specific input features (like individual pixels or words) carry the most "predictive information" for a neural network. By leveraging an information bottleneck approach directly in the input domain, InputIBA produces fine-grained, architecture-agnostic explanations that are highly faithful to the model's actual behavior.
The InputIBA Methodology
The core innovation of InputIBA is moving the information bottleneck from the hidden layers directly to the input domain. This allows the method to be model-agnostic (working for CNNs, RNNs, etc.) and pixel-perfect. The process operates in three main stages:
- Identify Deep Predictive Features: The method first finds an optimal bottleneck ($Z^*$) at a deep layer of the network. This step identifies the high-level, abstract concepts the model used to make its prediction.
- Generative Prior Mapping: Because mapping deep features directly back to specific pixels is mathematically challenging, the authors use a generative adversarial approach (minimizing Wasserstein distance) to find an input-level bottleneck ($Z_G$) that induces the same distribution as the deep features. This acts as a "prior," providing a general hint of where the predictive information lies.
- Input Bottleneck Optimization: Finally, the method optimizes a fine-grained mask ($\Lambda^*$) directly on the input. Guided by the prior from the previous step, this mask is constrained to keep only the specific input features that carry mutual information with the output.
Unlike previous methods that average across channels and lose resolution, InputIBA computes mutual information directly in the input space, allowing it to highlight specific edges, textures, or words rather than just broad, blurry regions.
Key Insights and Visual Evidence
The paper treats "predictive information" as the gold standard for feature importance: if removing a feature significantly reduces the mutual information between the input and the predicted class, that feature is objectively important.
This precision is best demonstrated by the "Broom vs. Trashcan" example. There is a known trade-off in explainable AI: smooth masks (like GradCAM) often perform better on certain metrics because they cover more area, but they lack precision. InputIBA provides the precision of gradient-based methods while maintaining the theoretical grounding of information theory.
As seen in the comparison above, standard IBA applied to early convolutional layers often assigns importance to irrelevant background objects (like the trashcan on the left side of the image). InputIBA (shown as $\Lambda^*$ on the far right) successfully ignores this background noise and focuses strictly on the object of interest—the broom—with remarkable, fine-grained fidelity.
To better quantify this localization capability, the authors propose a new metric called Effective Heat Ratios (EHR). EHR measures how well an attribution map aligns with ground-truth bounding boxes while accounting for the intensity of the "heat" (importance values). InputIBA scored 0.476 on this metric, significantly outperforming standard IBA (0.356) and GradCAM (0.283).
Experimental Results
The authors evaluated InputIBA across both Computer Vision (ImageNet using VGG16) and Natural Language Processing (IMDB sentiment analysis using a 4-layer LSTM) to prove its versatility.
- Sanity Checks: InputIBA successfully passed standard "Sanity Check" tests. Its explanations change significantly when model weights are randomized, proving it is explaining the model's learned behavior, not just acting as an edge detector for the input image.
- Superior Faithfulness (Vision): On ImageNet, InputIBA outperformed established baselines like DeepSHAP, Integrated Gradients, and GradCAM in rigorous faithfulness tests, including Sensitivity-N and Insertion/Deletion curves.
- NLP Success: Standard IBA struggles with Recurrent Neural Networks like LSTMs, often failing to provide meaningful explanations by assigning high importance to almost every word. InputIBA, however, successfully highlighted specific sentiment-carrying words (e.g., "best," "solid," "great") in the IMDB task, proving its architecture-agnostic claims.
Limitations and Future Directions
While highly accurate, InputIBA's primary limitation is computational complexity. Because it requires an additional optimization step for the generative mapping, it is slower than "one-pass" methods like GradCAM. Generating a single attribution map for a 224x224 image takes approximately 24 seconds on an RTX 3090 GPU. Additionally, like many bottleneck methods, the results can be sensitive to the choice of hyperparameter $\beta$, which controls the "tightness" of the bottleneck.
Despite the slower computation time, the authors suggest that InputIBA's improved precision makes it a strong candidate for mission-critical applications—such as medical imaging or autonomous driving—where "blurry" explanations are insufficient for establishing clinical or operational trust.
Summary updated