INAM: Image-Scale Neural Additive Models

Jana Hüls, Jan-Ole Perschewski, Sebastian Stober2025 · DOI

TL;DR

INAM (Image-Scale Neural Additive Models) makes image classification interpretable by combining a single layer of "human-readable" convolutional kernels with Neural Additive Models. While it matches state-of-the-art performance on simple datasets like MNIST and medical scans, it reveals a significant "interpretability tax" when faced with the high variability of complex real-world images.

What Problem Does This Paper Address?

Deep learning has a "black box" problem. While models like ResNet are incredibly accurate, it is often impossible to tell why they made a specific decision. In contrast, Neural Additive Models (NAMs) are inherently interpretable because they learn a separate, visualizable function for every input feature.

However, NAMs were designed for tabular data (like spreadsheets), where features like "Age" or "Income" have clear meanings. If you apply a NAM directly to an image, the "features" are individual pixels. Knowing that "Pixel (32, 45) being grey" contributed to a prediction doesn't help a human understand the model. The authors of this paper bridge this gap by creating INAM, which extracts meaningful visual features before passing them into an additive model.

Methods

The INAM architecture consists of three main stages:

  1. Interpretable Feature Extraction: The model uses a single convolutional layer with $N$ kernels. Unlike standard deep networks with dozens of layers, this single layer ensures that the features being detected are "shallow" and visually recognizable (e.g., an arch, a vertical line, or a specific texture).
  2. Global Max-Pooling: For each kernel, the model identifies only the maximum activation value. This effectively asks the question: "Is this specific shape present anywhere in the image?"
  3. Neural Additive Model (NAM): Each of these maximum activations is fed into its own small, dedicated neural network. The outputs of these networks are then summed to produce the final classification.

To ensure the kernels remain "interpretable," the authors use two specific regularization techniques:

  • Total Variation (TV) Loss: This penalizes "noisy" kernels, forcing the model to learn smoother, coarser structures that are easier for humans to identify.
  • Orthonormal Regularization: This encourages the kernels to be independent of one another, preventing the model from learning redundant features.

Datasets

The authors evaluated INAM on three distinct datasets to test different levels of visual complexity:

  • MNIST: The classic dataset of handwritten digits (0-9).
  • CRC (Colorectal Cancer): A histological dataset of tumor tissue slides, representing scientific data with relatively low variability.
  • CIFAR-10: A dataset of 10 classes of real-world objects (planes, dogs, etc.) which features high variability in color, lighting, and shape.

Key Insights and Ideas

  • The Interpretability-Complexity Trade-off: INAM performs exceptionally well on datasets with "low variability" (MNIST and CRC). When the visual patterns are consistent, a single layer of kernels is enough to capture the essence of the data.
  • Global vs. Local Explanations: Because of the additive structure, the authors can provide Global Explanations (showing how much a specific feature, like an "arch shape," contributes to the "digit 2" class across the whole dataset) and Local Explanations (creating "Decision Maps" that highlight exactly which parts of a specific image triggered the classification).
  • Feature Independence: By regularizing the kernels to be orthonormal, the authors ensure that the model doesn't "cheat" by spreading a single concept across multiple overlapping kernels, which would make the additive explanation harder to follow.

Key Results

The performance of INAM varies wildly depending on the complexity of the task:

Data Set Method Accuracy
MNIST INAM 0.9920
Linear Model 0.9385
CRC RBF Model 0.8740
INAM 0.8725
CIFAR-10 ResNet110 0.9357
INAM 0.6018

On MNIST, INAM is highly competitive. On the medical CRC dataset, it performs on par with existing interpretable methods. However, on CIFAR-10, accuracy drops to ~60% compared to ResNet's ~93%. This suggests that for complex, high-variability images, a single layer of interpretable features is currently insufficient to reach state-of-the-art performance.

Limitations and Future Work

The primary limitation is scalability. INAM struggles with high-variability datasets like CIFAR-10 because its "shallow" feature extraction cannot capture the hierarchical complexity (parts of objects, then objects, then scenes) that deep CNNs handle so well.

The authors suggest that while INAM is a promising step for explainable AI in fields like medicine or simple OCR, future work is needed to improve feature extraction without sacrificing the "additive" nature that makes the model so easy to interpret.

Summary updated

Abstract

Neural Additive Models are inherently interpretable models that can be applied to tabular data.However, when applying these models to images the value of a given pixel is not a meaningful feature for understanding the model.For that reason, we propose INAM -Image scale Neural Additive Model -a combination of trainable feature extractors and NAMs.We show INAMs can be successfully applied to image data sets with low variability while allowing global explanations of the models and data point-specific explanations.
BibTeX
@misc{huls2025inam,
  author = {Jana Hüls and Jan-Ole Perschewski and Sebastian Stober},
  title = {{INAM: Image-Scale Neural Additive Models}},
  year = {2025},
  doi = {10.14428/esann/2025.es2025-54},
}
Figure