• Transactions on Semiconductor Engineering
• /
• v.1 no.1
• /
• pp.23-30
• /
• 2023

With the recent development of Generative AI technology by IT giants, the size of the transformer model is increasing exponentially over trillion won. In order to continuously enable these AI services, it is essential to reduce the weight of the model. In this paper, we find a hardware-friendly structured pruning pattern and propose a lightweight method of the transformer model. Since compression proceeds by utilizing the characteristics of the model algorithm, the size of the model can be reduced and performance can be maintained as much as possible. Experiments show that the structured pruning proposed when pruning GPT-2 and BERT language models shows almost similar performance to fine-grained pruning even in highly sparse regions. This approach reduces model parameters by 80% and allows hardware acceleration in structured form with 0.003% accuracy loss compared to fine-tuned pruning.

• International Journal of Control, Automation, and Systems
• /
• v.1 no.3
• /
• pp.271-281
• /
• 2003

In this paper, a new nonlinear programming approach is suggested to solve biaffine matrix inequality (BMI) problems in multiobjective and structured controls. It is shown that these BMI problems are reduced to nonlinear minimization problems. An algorithm that is easily implemented with existing convex optimization codes is presented for the nonlinear minimization problem. The efficiency of the proposed algorithm is illustrated by numerical examples.

• The Journal of Korea Robotics Society
• /
• v.4 no.2
• /
• pp.147-154
• /
• 2009

Visions of ubiquitous robotics and ambient intelligence involve distributing information, knowledge, computation over a wide range of servers and data storage devices located all over the world, and integrating tiny microprocessors, actuators, and sensors into everyday objects as well in order to make them smart. In this paper, we introduce our ongoing research effort aimed at realizing ubiquitous robots in an information structured space. For this, a ubiquitous space and ambient intelligent systems for a librarian robot are introduced and the RFID technology based approach for these systems is described.

• Proceedings of the IEEK Conference
• /
• 2000.11d
• /
• pp.235-238
• /
• 2000

This paper describes the implementation of document image restoration system for the geometric distortion using structured light. To get accurate document images, the bounded book must be flattened by pushing down the book with a class plate. However, most of ancient documents are too fragile to be pushed. The proposed system restores the distorted character image due to geometric distortion.

• 한국신재생에너지학회:학술대회논문집
• /
• 2007.06a
• /
• pp.33-36
• /
• 2007

Currently, many structured catalysts using microchannel are researched to apply to fuel reforming. In this paper, ceramic monolith and metal mesh as structured catalysts are investigated for catalytic autothermal reforming. When GHSV increases, each structured catalyst has better performances(hydrogen production, fuel conversion) than packed bed catalyst for autothermal reforming. The major causes seem to be the elevated heat and mass transfer, gas phase reaction and redistribution of packed bed due to high pressure drop.

• 제어로봇시스템학회:학술대회논문집
• /
• 1995.10a
• /
• pp.420-423
• /
• 1995

The stability of system is one of the important aspects and to judge system's stability is another complicated problem. Previously, new technique derived from relaxing Lyapunov conditions has been already introduced and in this paper, this proposed technique applies to the practical dynamic systems. This utility of numerical procedures prove the comparable improvements of the estimation of robustness for dynamic systems having structured (bounded) perturbations.

• Journal of Institute of Control, Robotics and Systems
• /
• v.21 no.3
• /
• pp.233-236
• /
• 2015

In this paper, a nonlinear optimization problem is proposed to obtain a structured static output feedback controller for discrete time linear systems. The proposed optimization problem has LMI (Linear Matrix Inequality) constraints and a non-convex objective function. Using the conditional gradient method, we can obtain suboptimal solutions of the proposed optimization problem. Numerical examples show the effectives of the proposed approach.

• Journal of applied mathematics & informatics
• /
• v.18 no.1_2
• /
• pp.621-637
• /
• 2005

A hybrid method solving regularized structured linear total least norm (RSTLN) problems, which have highly ill-conditioned coefficient matrix with special structures, is suggested and analyzed. This scheme combining RSTLN algorithm and separation by parts guarantees the convergence of parameters and has an advantages in reducing the residual norm and relative error of solutions. Computational tests for problems arisen in signal processing and image formation process confirm that the presenting method is effective for more accurate solutions to (R)STLN problem than the (R)STLN algorithm.

• Proceedings of the Korean Information Science Society Conference
• /
• 2001.04b
• /
• pp.487-489
• /
• 2001

제본된 책이나 고문서를 디자타이즈할 때 책을 바르게 펴지 못하거나 지정된 위치나 방위로 위치시키지 못해서 문제가 발생한다. 특히 힘을 가할 수 없는 고문서의 경우에는 바르게 펼수가 없기 때문에 깊이 차이에 의한 문서의 왜곡이 발생한다. 본 연구에서는 Structured Light을 이용하여 깊이 차에 의한 문자의 왜곡을 복원하는 방안을 제안한다. 또한 입력시 책의 위치와 방위에 대한 제한을 완화시킬 수 있도록 책의 위치와 방위 변화에 적을 알 수 있는 방안을 제안한다,

• 제어로봇시스템학회:학술대회논문집
• /
• 2000.10a
• /
• pp.270-270
• /
• 2000

Input delay is frequently encountered in the practical systems since measurement delay and computational delay can be represented by input delay. In this viewpoint, this paper deals with the robust control problem of input delayed systems with structured uncertainty. Robust stability conditions are provided in terms of linear matrix inequalities(LMIs) and it is shown that the proposed conditions can give less conservative maximum bound of input delay guaranteeing robust stability.