COMPLEMENTARITY-BASED MULTI-LEVEL DISCRIMINATIVE FEATURE SPACE CONSTRUCTION
Ключевые слова:
deep learning, multi-level feature representation, discriminative learning, complementary features, orthogonal projection, feature redundancy, hierarchical representation, neural networks.Аннотация
Conventional multi-level feature fusion mechanisms in deep neural networks often fail to explicitly separate novel, complementary information from features already formed at preceding layers, introducing representational redundancy. This paper proposes a theoretical framework for Complementarity-Guided MultiLevel Discriminative Representation Learning (CLCD-RL). Representations extracted across network levels are mapped into a common space and decomposed into retained and complementary components via an orthogonal subspace projection mechanism. A task-aware gating function regulates residual contributions, while a supervised discriminative objective enforces intra-class compactness and inter-class separability. The theoretical and algebraic formulation establishes key conditions under which complementary features enhance discriminative power, providing a solid analytical foundation for hierarchical feature representations.
