SymbolNet: Bridging Latent Neural Representations and Symbolic Reasoning via Intermediate Feature Interpretation

The interpretation of intermediate representations in deep neural networks is critical for enhancing the transparency, trustworthiness, and applicability of artificial intelligence (AI) systems. In this paper, we propose SymbolNet, a framework that extracts mid-level features from trained models and transforms them into human-interpretable symbolic representations. SymbolNet constructs a symbolic graph composed of nodes and edges that capture both the semantic meaning and relational structure within the model’s internal reasoning process. This symbolic decoding bridges the model’s internal computations with human cognitive understanding, enabling structured and meaningful interpretation of AI behavior. Experimental results on the GTSRB dataset demonstrate that SymbolNet improves classification accuracy by 4% over the baseline and significantly enhances robustness against various noise conditions and adversarial attacks. Our work contributes to the field of explainable AI by introducing a novel approach that reveals the internal learning dynamics of non-interpretable models through symbolic reasoning.