Device for searching for squash function of capsule network and method for same

The squash function search apparatus using the AutoCAPS-ZERO framework and NAS principle optimizes squash functions in capsule networks, reducing computation time and maintaining accuracy, thereby addressing the high hardware costs of existing capsule network models.

WO2025058153A1 Publication Date: 2025-03-20EWHA UNIV IND COLLABORATION FOUND
2 Cites 0 Cited by

Patent Information

Application Number
PCT/KR2024/003896
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2023-09-13
Filing Date
2024-03-28
Publication Date
2025-03-20

AI Technical Summary

Technical Problem

Capsule network models, which offer improved accuracy through rotational immutability and spatial perception, incur high hardware costs due to increased computation complexity, particularly because the squash function consumes the largest number of operations.

Method used

The development of a squash function search apparatus and method using the AutoCAPS-ZERO framework based on the NEURAL Architecture Search (NAS) principle, which generates and evaluates initial squash functions to select optimal candidates that reduce computation time while maintaining accuracy.

Benefits of technology

This approach significantly reduces computation time while minimizing a decrease in accuracy by deriving optimal squash functions through the AutoCAPS-ZERO framework, thus addressing the high hardware costs associated with capsule network models.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The present invention relates to a device for searching for a squash function and a method for same. The device for searching for a squash function according to an embodiment comprises: a data reception unit that receives at least one of a capsule network model, a target task, or a capsule network-related hyperparameter as input data; a function generation unit that generates a plurality of initial squash functions on the basis of the input data; a performance evaluation unit that uses a fitness function based on Kullback-Leibler divergence to evaluate the performance of the plurality of initial squash functions; and a function search unit that selects at least one candidate squash function among the plurality of initial squash functions on the basis of evaluation results for the plurality of initial squash functions, and generates a search result based on the selected candidate squash function.
Need to check novelty before this filing date? Find Prior Art