Instance-Adaptive Gated Fusion Of Multi-Transform Image Representations

This dissertation proposes the Instance-Adaptive Gated Fusion (IAGF) framework, a novel deep learning architecture for adaptive and interpretable fusion of multiple time–series image transformations. While existing methods rely on static concatenation or dataset-level optimization, IAGF introduces a learnable gating mechanism that dynamically assigns per-instance weights to Recurrence Plots (RP), Gramian Angular Summation Fields (GASF), and Gramian Angular Difference Fields (GADF). The gating layer performs a convex fusion of transformation-specific embeddings under a softmax constraint, ensuring mathematical stability and interpretability. An entropy-regularized objective prevents dominance collapse and promotes balanced exploration of transformations during training. Comprehensive experiments across eighteen benchmark datasets, spanning univariate, multivariate, and tabular modalities, demonstrate that IAGF consistently outperforms both single-transform CNNs and static fusion baselines. Notably, the model achieves substantial gains on complex datasets such as Adiac, ChlorineConcentration, and ArticularyWordRecognition, while maintaining robust performance on non–time–series domains. Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations confirm that gating weights align with spatially discriminative regions, offering intrinsic interpretability without post hoc analysis. Theoretically, IAGF is supported by a rigorous convexity proof and bounded stability analysis presented in the Appendix. Its modular design generalizes across domains, paving the way for applications in multimodal, biomedical, and environmental data contexts. Overall, this work establishes IAGF as a principled, explainable, and domain-agnostic framework for adaptive representation fusion, advancing the state of transformation-based time–series learning and interpretable deep models.