SINF intelligently clusters inputs to selectively engage DNN subgraphs, reducing computational load while maintaining high accuracy.
Copy Image URL via instructions below
https://nu.testtechnologypublisher.com/files/sites/mark-saulich-21.jpg
Background
The deployment of Deep Neural Networks in real-world applications faces significant challenges due to their substantial computational and data labeling requirements. Traditional techniques aimed at optimizing these networks for resource-constrained environments, such as pruning and quantization, often result in a notable decrease in accuracy.
Description
Northeastern researchers have created a novel framework known as Semantic Inference (SINF) that revolutionizes the optimization process of Deep Neural Networks (DNNs). This technique intelligently clusters semantically similar inputs, activating only the necessary portions of the network for processing. This selective engagement dramatically reduces the computational load while maintaining high accuracy levels. The core innovation lies in a supplementary classifier that swiftly identifies the semantic cluster of incoming data, followed by the engagement of a specific subgraph of the base Deep Neural Network (DNN) tailored for that cluster. To identify these critical subgraphs efficiently, the team has introduced a new approach named Discriminative Capability Score (DCS). This methodology ensures that only the most relevant and efficient pathways within the network are utilized during inference, striking an optimal balance between performance and computational efficiency.
Benefits
Applications
Opportunity