Entanglement Learning: an Information-Theoretic Framework for Adaptive Convolutional Neural Networks

While Convolutional Neural Networks (CNNs) have transformed industrial computer vision systems, their adaptability to dynamic environments post-deployment remains a critical challenge, particularly when encountering out-ofdistribution objects, or unseen environments. We introduce Entanglement Learning (EL), an information-theoretic framework that enables CNN adaptability by measuring the mutual information between multiple layers (e.g. convolutional layer activations, fully connected layer output, and output class distributions). Our framework quantifies these relationships in bits, providing metrics for detecting environmental changes such as sensor drift, hardware degradation, or feature distribution shifts. Out-of-distribution samples and adversarial perturbations generate distinctive entropy patterns that diverge from the training distribution's information flow. We implement this framework through an Information Digital Twin (IDT), a computational module operating parallel to the CNN that establishes a performance baseline by analyzing CNN mutual information patterns under nominal conditions, then monitors variations from baseline to generate parameter adjustment signals. This concept can be applied to industrial vision systems for autonomous vehicles adapting to unseen environmental conditions and robotic vision systems enabling precise object manipulation. This work establishes a conceptual foundation for developing self-monitoring computer vision systems capable of maintaining robust performance while providing inherent safeguards against distribution shifts and adversarial manipulation.