Training Resilient AI Models With Rich Interpretations From Highly Scarce Data

Traditional neural networks struggle when trained on limited data and lack inherent interpretability. This limitation arises because conventional NNs operate as tabula rasa, relying entirely on the volume and quality of training data for learning, without any inherent knowledge or structure to guide them. Inspired by natural neural networks in living beings, which exhibit innate intelligence through purpose-driven design even before learning begins, we propose a novel framework, WINET (White box Interpretable Neural Networks). Unlike traditional neural networks, WINET integrates domain knowledge directly into the architecture from the outset, enabling relatively robust learning even with minimal training data. Our experiments reveal that, much like human learning, WINET requires significantly less training data than traditional neural networks, while maintaining resilience against training data scarcity. Additionally, WINET enhances interpretability – an essential attribute for AI models involved in critical decision making — where conventional neural networks often fall short. To validate WINET’s effectiveness, we first compare it qualitatively with existing interpretable models, then quantitatively apply it to predicting mobile network coverage, a complex task influenced by both controlled and random variables. We compared WINET against two common alternatives to system modeling – (i) analytical models and (ii) conventional AI (black box neural networks) – using both simulated and real data. Results show that conventional AI exhibits a drastic performance drop with scarce data (realistic drive test data), with Mean Squared Error increasing by 2200%. The analytical model also performs poorly. In contrast, WINET shows superior generalization and resilience to limited training data in unseen test scenarios.