Additional experiments

PermGrad additional experiments and revision updates

This release provides the updated experimental and reproducibility materials associated with the revised version of the manuscript:

PermGrad: Interpretable Hybrid Neural Networks with Synthetic Images for Tabular Data

PermGrad is an interpretability framework for Hybrid Neural Networks (HyNNs) that jointly process original tabular features and TINTO-generated synthetic images. The architecture combines an MLP branch, which captures symbolic/global dependencies in the original tabular space, with a CNN branch, which learns spatially organized feature interactions from synthetic image representations. The proposed PermGrad mechanism integrates permutation-based feature importance, Grad-CAM saliency, and pruning-based branch contribution weights into a unified, modality-aware global feature attribution score.

Main updates in this release

• Added additional resolution ablation experiments for TINTO-generated synthetic images using 10×10, 20×20, and 30×30 image sizes.
• Included quantitative performance comparisons across image resolutions for representative classification and regression datasets.
• Extended the class-imbalance robustness analysis to both GAS and Covertype datasets.
• Added Spearman correlation analyses between class size and Grad-CAM saliency/entropy to assess whether PermGrad exhibits majority-class saliency bias.
• Added feature-ranking stability analysis under class-balanced retraining.
• Updated interpretability analysis to distinguish between moderate imbalance and extreme imbalance scenarios.
• Improved the visual structure and readability of the HyNN–PermGrad workflow figure.
• Revised terminology related to synthetic image interpretability, replacing claims of exact 1-to-1 feature–pixel mapping with more precise notions of deterministic feature localization and pixel-level traceability.
• Updated manuscript-related materials, figures, and tables to reflect the latest revision.

Repository contents

This repository includes the code, notebooks, figures, and reproducibility resources required to reproduce the experiments reported in the manuscript, including:

• TINTO-based tabular-to-image generation.
• MLP, CNN, and HyNN model training.
• Permutation-based feature importance for the MLP branch.
• Grad-CAM saliency extraction for the CNN branch.
• PermGrad branch-weighted feature attribution.
• Resolution ablation experiments.
• Class-imbalance robustness analysis.
• Additional interpretability and visualization scripts.

Notes

This release is intended to preserve the experimental state associated with the revised manuscript submission. The corresponding Zenodo archive provides a citable snapshot of the repository.

v1.1.0: SOTA Baselines, Training Dynamics, and Scalability Optimizations

Overview

This major release incorporates extensive new experiments, baselines, and scalability optimizations, developed in response to rigorous peer-review feedback. The HyNN-PermGrad framework has been substantially expanded to include modern deep tabular baselines and detailed training dynamics analyses.

🚀 New Features & Experiments

• SOTA Deep Tabular Baselines: Added training and evaluation scripts for modern architectures, specifically FT-Transformer and TabNet, allowing for direct comparison against the HyNN architecture.
• Training Dynamics & Gradient Tracking: Introduced new logging mechanisms to extract and visualize MSE loss curves and Gradient L2 Norms during training. This allows for the empirical demonstration of the MLP branch acting as a "symbolic anchor" to stabilize gradient flow.
• Resolution Ablation (TINTOlib): Added scripts to generate and evaluate synthetic images at multiple resolutions (10x10, 20x20, 30x30) to analyze topological balance, pixel collisions, and spatial sparsity.
• Adaptive Permutation Sampling: Implemented a highly scalable, two-stage permutation sampling strategy for high-dimensional datasets (e.g., GAS). This optimization reduces the computational overhead by >50% while preserving exact feature rankings.
• Class Imbalance Analysis: Added evaluation modules to assess PermGrad's robustness and spatial attention entropy under severe class imbalance scenarios.

🛠️ Code Improvements & Refactoring

• Enforced unblurred configurations by default in TINTOlib generation scripts to guarantee exact 1-to-1 feature-pixel mapping for interpretability.
• Cleaned up the static late-fusion architecture code for better reproducibility and isolation of branch-wise contributions.
• Updated requirements.txt to include new dependencies required for the SOTA baselines.

📊 Reproducibility

All metrics, weights, and logs associated with the revised manuscript's benchmark (HELOC, Dengue, Covertype, GAS, Puma) can be reproduced using the updated benchmark scripts included in this release.

PermGrad v1.0.0 – Official Release

This release contains the full code, experiments, synthetic-image generation pipelines, and interpretability analyses used in the manuscript:

"PermGrad: A Hybrid Neural Network Framework for Unified Symbolic–Spatial Interpretability in Tabular Deep Learning"

Included in this release:

• Training scripts for MLP, CNN, and HyNN architectures (classification and regression)
• TINTO-based synthetic image generation workflow
• Permutation importance, Grad-CAM, and PermGrad interpretability modules
• All experiment configurations and reproducible settings
• Graphical abstract and figures used in the paper

This version corresponds to the exact codebase used for the submission to Pattern Recognition – Special Issue on Evolving Multi-View Learning.