• Title/Summary/Keyword: Fine Tuning

Search Result 343, Processing Time 0.034 seconds

Building robust Korean speech recognition model by fine-tuning large pretrained model (대형 사전훈련 모델의 파인튜닝을 통한 강건한 한국어 음성인식 모델 구축)

  • Changhan Oh;Cheongbin Kim;Kiyoung Park
    • Phonetics and Speech Sciences
    • /
    • v.15 no.3
    • /
    • pp.75-82
    • /
    • 2023
  • Automatic speech recognition (ASR) has been revolutionized with deep learning-based approaches, among which self-supervised learning methods have proven to be particularly effective. In this study, we aim to enhance the performance of OpenAI's Whisper model, a multilingual ASR system on the Korean language. Whisper was pretrained on a large corpus (around 680,000 hours) of web speech data and has demonstrated strong recognition performance for major languages. However, it faces challenges in recognizing languages such as Korean, which is not major language while training. We address this issue by fine-tuning the Whisper model with an additional dataset comprising about 1,000 hours of Korean speech. We also compare its performance against a Transformer model that was trained from scratch using the same dataset. Our results indicate that fine-tuning the Whisper model significantly improved its Korean speech recognition capabilities in terms of character error rate (CER). Specifically, the performance improved with increasing model size. However, the Whisper model's performance on English deteriorated post fine-tuning, emphasizing the need for further research to develop robust multilingual models. Our study demonstrates the potential of utilizing a fine-tuned Whisper model for Korean ASR applications. Future work will focus on multilingual recognition and optimization for real-time inference.

Neural Network Method for Tuning PID Gains (신경회로망을 이용한 PID 제어기의 이득조정)

  • Moon, Seok-Woo;Lee, Chong-Ho
    • Proceedings of the KIEE Conference
    • /
    • 1992.07a
    • /
    • pp.476-479
    • /
    • 1992
  • This paper presents a neural network method for tuning PlD controller of a time-varying process. Three gains of PlD controller are tuned for a certain desirable response pattern by back-propagation neural network. The neural network is trained using changes of output features vs. changes of PlD gains. But sometimes it needs longer training time and larger structure to train the correlation between the process and controller on entire region of the process. The difficulty in system identification is that the inverse function of the system can not be clearly stated. To cope with the problem, we do not train the neural network to respond correctly for the entire regions but train for only local region where the system is heading toward by training the neural network and tuning of the PlD controller. It may be trained for fine-tuning itself. Simulation results show that the adaptive PID controller using neural network trained in the local area performs remarkably for time-varying second order process.

  • PDF

A Fuzzy Self-Tuning PID Controller with a Derivative Filter for Power Control in Induction Heating Systems

  • Chakrabarti, Arijit;Chakraborty, Avijit;Sadhu, Pradip Kumar
    • Journal of Power Electronics
    • /
    • v.17 no.6
    • /
    • pp.1577-1586
    • /
    • 2017
  • The Proportional-Integral-Derivative (PID) controller is still the most widespread control strategy in the industry. PID controllers have gained popularity due to their simplicity, better control performance and excellent robustness to uncertainties. This paper presents the optimal tuning of a PID controller for domestic induction heating systems with a series resonant inverter for controlling the induction heating power. The objective is to design a stable and superior control system by tuning the PID controller with a derivative filter (PIDF) through Fuzzy logic. The paper also compares the performance of the Fuzzy PIDF controller with that of a Ziegler-Nichols PID controller and a fine-tuned PID controller with a derivative filter. The system modeling and controllers are simulated in MATLAB/SIMULINK. The results obtained show the effectiveness and superiority of the proposed Fuzzy PID controller with a derivative filter.

A 2.4 ㎓ Back-gate Tuned VCO with Digital/Analog Tuning Inputs (디지털/아날로그 입력을 통해 백게이트 튜닝을 이용한 2.4 ㎓ 전압 제어 발진기의 설계)

  • Oh, Beom-Seok;Hwang, Young-Seung;Chae, Yong-Doo;Lee, Dae-Hee;Jung, Wung
    • Proceedings of the IEEK Conference
    • /
    • 2003.11c
    • /
    • pp.32-36
    • /
    • 2003
  • In this work, we have designed a fully integrated 2.4GHz LC-tuned voltage-controlled oscillator (VCO) with multiple tuning inputs for a 0.25-$\mu\textrm{m}$ standard CMOS process. The design of voltage-controlled oscillator is based on an LC-resonator with a spiral inductor of octagonal type and pMOS-varactors. Only two metal layer have been used in the designed inductor. The frequency tuning is achieved by using parallel pMOS transistors as varactors and back-gate tuned pMOS transistors in an active region. Coarse tuning is achieved by using 3-bit pMOS-varactors and fine tuning is performed by using back-gate tuned pMOS transistors in the active region. When 3-bit digital and analog inputs are applied to the designed circuits, voltage-controlled oscillator shows the tuning feature of frequency range between 2.3 GHz and 2.64 GHz. At the power supply voltage of 2.5 V, phase noise is -128dBc/Hz at 3MHz offset from the carrier. Total power dissipation is 7.5 mW.

  • PDF

Hybrid Genetic Algorithm Reinforced by Fuzzy Logic Controller (퍼지로직제어에 의해 강화된 혼합유전 알고리듬)

  • Yun, Young-Su
    • Journal of Korean Institute of Industrial Engineers
    • /
    • v.28 no.1
    • /
    • pp.76-86
    • /
    • 2002
  • In this paper, we suggest a hybrid genetic algorithm reinforced by a fuzzy logic controller (flc-HGA) to overcome weaknesses of conventional genetic algorithms: the problem of parameter fine-tuning, the lack of local search ability, and the convergence speed in searching process. In the proposed flc-HGA, a fuzzy logic controller is used to adaptively regulate the fine-tuning structure of genetic algorithm (GA) parameters and a local search technique is applied to find a better solution in GA loop. In numerical examples, we apply the proposed algorithm to a simple test problem and two complex combinatorial optimization problems. Experiment results show that the proposed algorithm outperforms conventional GAs and heuristics.

"차이를 만들지 않는다(Makes No Difference)"는 논증과 반사실적 조건문에 대한 분석

  • Kim, Sea-Hwa
    • Korean Journal of Logic
    • /
    • v.10 no.2
    • /
    • pp.1-22
    • /
    • 2007
  • 본 논문에서 필자는 먼저 "차이를 만들지 않는다(Makes No Difference)"는 논증(줄여서 MND 논증)을 소개하고 이에 대한 앨런 베이커(Alan Baker)의 반론을 자세히 살펴본다. MND 논증, 특히 그 전제를 제대로 평가하기 위해서는 반사실적 조건문에 대한 분석이 필요한데, 베이커는 이를 위해 이갈 크바크(Igal Kvart)의 분석을 이용한다. 이에 필자는 크바트의 분석을 비판함으로써 이에 의존한 베이커의 주장을 비판한다. 이를 위해 필자는 Fine-Tuning 논증을 예로 삼아, 크바트의 반사실적 조건문에 대한 분석, 특히 그의 반법칙적 조건문에 대한 분석은 실제 과학자들의 논의와도 상충되며, 이 과학자들의 논의를 바탕으로 한 철학적 논쟁과도 상충되기 때문에 그의 분석을 받아들일 수 없음을 보인다. 이를 통해 필자는 크바트의 분석에 의존하여 MND 논증을 공격한 베이커의 결론도 마찬가지로 받아들일 수 없다는 것을 보인다.

  • PDF

Performance Analysis of Feature Extractor for Transfer Learning of a Small Sample of Medical Images (소표본 의료 영상의 전이 학습을 위한 Feature Extractor 기법의 성능 비교 및 분석)

  • Lee, Dong-Ho;Hong, Dae-Yong;Lee, Yeon;Shin, Byeong-Seok
    • Proceedings of the Korea Information Processing Society Conference
    • /
    • 2018.05a
    • /
    • pp.405-406
    • /
    • 2018
  • 본 논문은 소표본 의료용 영상 분석의 정확도 향상을 위해 전이학습 모델을 feature extractor로 구축하여 학습시키는 방법을 연구하였으며 성능 평가를 위해 선학습모델로 AlexNet, ResNet, DenseNet을 사용하여 fine tuning 기법을 적용하였을 때와의 성능을 비교 분석하였다. 그 결과 실험에 사용된 3개의 모델에서 fine tuning 기법보다 향상된 정확도를 보임을 확인하였고, 또한 ImageNet으로 학습된 AlexNet, ResNet, DenseNet이 소표본 의료용 X-Ray 영상에 적용될 수 있음을 보였다.

GAN-based Color Palette Extraction System by Chroma Fine-tuning with Reinforcement Learning

  • Kim, Sanghyuk;Kang, Suk-Ju
    • Journal of Semiconductor Engineering
    • /
    • v.2 no.1
    • /
    • pp.125-129
    • /
    • 2021
  • As the interest of deep learning, techniques to control the color of images in image processing field are evolving together. However, there is no clear standard for color, and it is not easy to find a way to represent only the color itself like the color-palette. In this paper, we propose a novel color palette extraction system by chroma fine-tuning with reinforcement learning. It helps to recognize the color combination to represent an input image. First, we use RGBY images to create feature maps by transferring the backbone network with well-trained model-weight which is verified at super resolution convolutional neural networks. Second, feature maps are trained to 3 fully connected layers for the color-palette generation with a generative adversarial network (GAN). Third, we use the reinforcement learning method which only changes chroma information of the GAN-output by slightly moving each Y component of YCbCr color gamut of pixel values up and down. The proposed method outperforms existing color palette extraction methods as given the accuracy of 0.9140.

Analysis of Automatic Topic Classification using Youtube Meta Information (유튜브 메타정보를 이용한 자동 주제 분류 고찰)

  • Kim, Yong-Woo;Jeon, Seong-Bae;Jung, Yuchul
    • Proceedings of the Korean Society of Computer Information Conference
    • /
    • 2021.01a
    • /
    • pp.349-351
    • /
    • 2021
  • Youtube 동영상 업로드 시, 사용자가 직접 주제를 설정해야 하는 어려움이 있다. 본 연구에서는 사용자가 입력하는 제목과 설명정보를 이용하여 자동으로 주제를 분류하는 연구를 진행하였다. 이를 위해 한국어기반의 컨텐츠 중 고빈도의 8개 주제 카테고리를 선정하고, 이를 1.3만건의 학습데이터를 크롤링을 통해 구축하였다. 또한, 다양한 알고리즘들에 대한 최대성능을 확인하기 위해 대표적인 텍스트 분류 방법인 SVM과 LSTM기법 및 BERT 모델기반 미세적용(fine-tuning)을 시도하였다. 결과적으로 Bert-multiligual (base)를 fine-tuning한 실험에서 최대 94%의 정확도를 확인하였다. 하지만, Youtube 동영상 특성상 여러 주제를 가진 것들이 상당수 존재하기에, 실제 체감정확도는 더 높을 것으로 기대된다.

  • PDF

Metonymy Resolution based on Neural Approach (딥러닝 방식을 이용한 환유 해소)

  • Whang, Taesun;Lee, Chanhee;Yang, Kisu;Lee, Dongyub;Koo, Youngeun;Jeon, Taehee;Lim, Heuiseok
    • Annual Conference on Human and Language Technology
    • /
    • 2019.10a
    • /
    • pp.375-379
    • /
    • 2019
  • 언어학에서의 환유법은 표현을 위해 빌려온 대상이 다양한 의미로 해석 가능하기에 매우 어렵고 난해한 분야이다. 환유의 특성 상 주어진 엔티티의 환유 여부를 구분하기 위해서는 앞뒤 단어와의 연관성 뿐만 아니라 문장 전체의 문맥 정보에 대한 고려가 필수적이다. 최근 이러한 문맥 정보를 고려하여 학습된 다양한 모델들이 등장하면서 환유법에 대한 연구를 하기에 좋은 환경이 구축되고 있다. 본 논문에서는 언어학적 자질 정보를 최소화한 딥러닝을 이용한 환유 해소 모델을 제안한다. LSTM 기반의 feature-based 모델과 및 BERT, XLNet, RoBERTa와 같은 fine-tuning 모델들에 대한 실험을 진행하였다. 실험 결과, fine-tuning 모델들이 baseline과 비교하여 뛰어난 성능 향상을 가져왔으며, 특히 XLNet 모델은 두 개의 환유 해소 데이터 SemEval 2007와 ReLocaR에 대해 각각 90.1%과 95.8%의 정확도를 보여주었다.

  • PDF