Neural Additive Models

NAM (Neural Additive Models) are interpretable neural networks that learn a separate shape function for each input feature — $f(x) = \beta_0 + \sum_i f_i(x_i)$, where each $f_i$ is a small neural network, providing the interpretability of GAMs with the flexibility of neural networks.

How NAMs Work

• Feature Networks: Each input feature $x_i$ has its own small neural network $f_i$ that outputs a scalar.
• Addition: The final prediction is the sum of all feature contributions: $f(x) = \beta_0 + \sum_i f_i(x_i)$.
• Visualization: Each $f_i(x_i)$ can be plotted as a shape function — showing the effect of each feature.
• Training: Standard backpropagation with dropout and weight decay for regularization.

Why It Matters

• Interpretable: The contribution of each feature is independently visualizable — no interaction hiding effects.
• Non-Linear: Unlike linear models, each $f_i$ can capture arbitrary non-linear effects.
• Glass-Box: NAMs provide "glass-box" interpretability comparable to linear models with much better accuracy.

NAMs are interpretable neural nets by design — isolating each feature's contribution through separate sub-networks for transparent predictions.