Introduction
Artificial neural networks (ANNs) are powerful function approximators but are often criticized for being "black boxes" due to their extreme over-parameterization and lack of transparency. This paper introduces input-skip Latent Binary Bayesian Neural Networks (ISLaB), a novel architecture designed to address these issues. By allowing inputs to bypass hidden layers and route directly to the output, ISLaB aggressively prunes redundant weights, resulting in a highly sparse, interpretable model with built-in global and local explanations.
The Challenge of Over-parameterization
Standard ANNs force input data through every hidden layer, regardless of the complexity of the underlying relationship. If a feature's relationship with the output is purely linear, passing it through multiple non-linear layers is inefficient and obscures the model's logic. While Bayesian Neural Networks (BNNs) provide a framework for uncertainty quantification by treating weights as probability distributions rather than point estimates, they typically still follow fixed, dense architectures.
Furthermore, explaining complex models often relies on post-hoc tools like SHAP or LIME, which are themselves approximations and may not accurately reflect the model's internal state or uncertainty. The authors set out to build a network that learns its own optimal structure and explains itself out-of-the-box.
The ISLaB Methodology
The authors extend Latent Binary Bayesian Neural Networks (LBBNNs), which utilize a "spike-and-slab" prior to learn the probability of weight inclusion. Key innovations include:
- Input-Skip Connections: Original input covariates are concatenated to the input of every hidden layer and the output layer. This enables the network to learn the optimal routing for each feature, allowing linear signals to skip hidden layers entirely.
- Active Path Formalization: The model identifies "active paths"—sequences of non-zero weights from input to output. Nodes or layers without active paths are mathematically redundant and removed, leading to extreme sparsity.
- Exact Explainability: By utilizing piecewise-linear activation functions (e.g., ReLU), the network behaves as a Generalized Linear Model (GLM) for any specific input. This allows for the extraction of exact local explanations via gradients, complete with Bayesian credibility intervals.
Experimental Evaluation
The ISLaB approach was tested on a variety of datasets, ranging from synthetic linear/non-linear data to complex real-world benchmarks including the Wisconsin Breast Cancer dataset, Mice Protein expression, and image datasets like FER2013 and MNIST.
Key Results and Insights
- State-of-the-Art Compression: On MNIST, ISLaB achieved ~97% accuracy using only 935 weights—a reduction of over 99.9% from the original 1.3 million parameters. This outperforms other sparse modeling techniques like the Lottery Ticket Hypothesis.
- Adaptive Complexity: The model successfully recovers the underlying structure of data. For linear synthetic data, it prunes all hidden layers to become a simple linear regression.
- Superior Calibration: ISLaB consistently outperformed frequentist L1-regularized networks in Expected Calibration Error (ECE), indicating better-calibrated uncertainty.
- Intrinsic Interpretability: The model provides clear global explanations through its sparse connectivity and local explanations via gradient heatmaps that highlight relevant features (e.g., pixels around the mouth for "happy" face detection).
Limitations and Future Directions
While promising, the model's performance is sensitive to the initialization of weight inclusion probabilities. "Dense" initializations favor accuracy, while "sparse" ones favor compression. Additionally, the discrete nature of structure learning can pose challenges for standard gradient-based optimizers. Future work may explore adaptive initialization strategies or MCMC-based optimization to improve convergence and escape local minima.
Summary updated