Abstract
In modern deep learning architectures, computational efficiency is a primary bottleneck, particularly when deploying large-scale neural networks to edge devices or real-time systems. Dynamic Input-Correlated Layer Execution (DICLE) is proposed as a theoretical framework designed to optimize inference latency by dynamically bypassing or executing specific layers of a deep neural network based on the statistical correlation of input features.
By evaluating the spatial and temporal correlation of input representations at runtime, the DICLE framework determines whether down-stream layers will yield statistically significant updates to the activation state. If the correlation exceeds a predefined threshold, redundant layer computations are bypassed, and intermediate state estimates are projected forward, saving substantial floating-point operations (FLOPs) without compromising accuracy.
Key Contributions
- Input-Correlation Metric: Formulated a mathematical metric to evaluate feature correlation across layer interfaces dynamically at runtime.
- State Projection Mechanics: Developed a mathematical model to approximate outputs of skipped layers based on correlation profiles, preventing gradient disconnects and state corruption.
- Dynamic Routing Algorithm: Outlined an execution engine that adjusts threshold sensitivities adaptively, matching hardware thermal throttling or latency requirements.
Publication Details
- Journal: International Journal of Creative Research Thoughts (IJCRT)
- Paper ID: IJCRT26A4022
- Publication Date: April 2026
- Co-Authors: Tanish Jagtap, Vidish Bajpai