Inherently Interpretable Deep Learning via Neural Additive Models

65 concepts55 papers1 proposal

Deep neural networks are incredibly powerful, but their internal representations are notoriously opaque. In a standard multi-layer perceptron (MLP), input features are immediately mixed together in the very first hidden layer. As data flows deeper into the network, these features entangle in high-dimensional space, making it virtually impossible to isolate the exact impact of any single variable on the final prediction.

To understand these "black boxes," practitioners often turn to post-hoc explanation methods like SHAP or LIME. However, these tools only approximate a model's behavior locally and can be unfaithful to its true internal mechanics.

Neural Additive Models (NAMs) solve this problem by changing the architecture itself. Instead of allowing features to entangle, NAMs force the network to remain inherently interpretable by design. They do this by combining the structural transparency of classical Generalized Additive Models (GAMs) with the non-linear expressiveness of deep learning.

The Core Architecture: Feature-Wise Decomposition

The fundamental constraint of a Neural Additive Model is simple: prevent features from interacting within hidden layers.

Mathematically, a standard regression model maps an input vector $x = (x_1, x_2, \dots, x_d)$ to a target using a complex multivariate function $f(x)$. A Neural Additive Model decomposes this prediction into a sum of independent, univariate functions:

$$f(x) = \beta_0 + f_1(x_1) + f_2(x_2) + \dots + f_d(x_d)$$

To implement this structurally, the architecture assigns a dedicated, independent neural sub-network to each individual input feature.

x_1 ───► [ Sub-Network 1 ] ───► f_1(x_1) ──┐
x_2 ───► [ Sub-Network 2 ] ───► f_2(x_2) ──┼──► [ Summation ] ──► Prediction
x_3 ───► [ Sub-Network 3 ] ───► f_3(x_3) ──┘

As detailed by Rishabh Agarwal et al. in their seminal paper Neural Additive Models: Interpretable Machine Learning with Neural Nets, each sub-network takes only a single scalar input and outputs a single scalar contribution score. These individual scores are aggregated through a final summation layer to produce the global prediction.

Because the sub-networks never share weights or hidden states, the marginal effect of any given feature is completely isolated. The model cannot implicitly learn hidden cross-feature correlations.

Why It Works: Exact Shape Functions

The primary benefit of this feature-wise decomposition is that the model's exact internal logic can be plotted and audited. By passing a range of synthetic values through a single sub-network, practitioners can generate a continuous curve known as a shape function.

In Explainable Automatic Grading with Neural Additive Models, Aubrey Condor et al. demonstrate how these shape functions allow humans to verify that an automated system is rewarding the correct underlying concepts rather than exploiting spurious patterns. If a feature's shape function behaves erratically, it signals a flaw in the model's reasoning that can be immediately diagnosed.

This exact trace-ability makes NAMs highly valuable in high-stakes domains where black-box models pose safety or regulatory risks:

Retaining Expressiveness Without Entanglement

A common skepticism regarding additive models is that they sacrifice too much predictive power by ignoring the interactions between features. However, the individual sub-networks themselves are deep and highly expressive.

Federico Siciliano et al. prove in NEWRON: A New Generalization of the Artificial Neuron to Enhance the Interpretability of Neural Networks that architectures processing inputs independently before aggregation are still universal approximators of continuous multivariate additive functions. While traditional GAMs rely on rigid linear coefficients or greedy spline-fitting algorithms, the deep sub-networks in a NAM use gradient descent to learn highly complex, non-linear univariate shapes.

Capturing Controlled Interactions

To bridge the gap when pure additivity is insufficient, researchers introduce explicit, low-order interaction blocks into the summation layer.

Michael Tsang et al. address this in Neural Interaction Transparency (NIT): Disentangling Learned Interactions for Improved Interpretability by constraining networks to learn interactions within independent structural blocks. This allows the model to capture essential pairwise relationships without letting them cascade into uninterpretable multi-feature entanglements.

Similarly, Yi-Ting Hung et al. propose the Sparse Deep Additive Model with Interactions: Enhancing Interpretability and Predictability, which assigns specific feature pairs to their own dedicated deep sub-networks. The final prediction remains a transparent summation, but it now includes both univariate shape curves and bivariate shape heatmaps. As David Köhler demonstrates in Achieving interpretable machine learning by functional decomposition of black-box models into explainable predictor effects, this functional decomposition isolates main effects and pairwise interactions into separate, visualizable subfunctions, preserving global transparency.

Alternative Backbones

The core philosophy of neural additive modeling—isolating features into independent components before a final summation—is not restricted to standard multi-layer perceptrons. The sub-network backbone can be swapped to suit different data modalities or computational constraints.

For example, Wei Zhang 0262 et al. introduce Gaussian Process Neural Additive Models, replacing deep neural networks with Gaussian Processes approximated by random Fourier features to dramatically accelerate training speeds. Thummaluru Siddartha Reddy extends this paradigm to graph-structured data in Interpretable and Parameter Efficient Graph Neural Additive Models with Random Fourier Features.

To better handle categorical features in tabular datasets, Arik Reuter develops Interpretable Additive Tabular Transformer Networks, which routes categorical embeddings through a specialized transformer while maintaining the overall additive constraint. By summing the scalar outputs of these diverse backbones, the framework preserves its "glass-box" nature while leveraging modern architectural advancements.

Key Foundational Work

Correcting Forecasts with Multifactor Neural Attention. (2016) — contributed 1 concept elements to this direction.
Additive Kernels for High-dimensional Gaussian Process Modeling (2010) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_additive_neural_architecture_ef087394,
  title = {additive neural architecture},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/additive-neural-architecture/rdr_ef08739479f84dba983d8c9f93146847},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_additive_neural_architecture_ef087394_v1,
  title = {additive neural architecture},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/additive-neural-architecture/rdr_ef08739479f84dba983d8c9f93146847/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake