Earticle

In this paper, we propose a PTQ (Post Training Quantization) static with QO (Quantization Only) technique to efficiently deploy and execute deep learning models. A comparative performance evaluation of PTQ static with QDQ (Quantize and DeQuantize) and the proposed quantization method was conducted using the MNIST (Modified National Institute of Standards and Technology database) dataset and 8- bit quantization. Experimental results indicate that the PTQ static with QO method reduces the size of the model by approximately 33%, increases the inference speed by 1.5 times, and minimizes the accuracy loss, similar to the PTQ static with QDQ method. The proposed PTQ static with QO method offers a significant technical enhancement to facilitate the efficient deployment and execution of AI (Artificial Intelligence) models through the quantization of deep learning models. We have shown that the PTQ static with QO method is a beneficial and efficient approach to decrease the size and computation of deep learning models. This study makes novel contributions to the quantization of deep learning models. The practical potential of the PTQ static with QO method lies in its ability to be more suitably deployed for the purposes of AI hardware.

The rising popularity of intelligent embedded systems, coupled with the substantial computational and memory requirements of convolutional neural networks (CNNs), necessitates cost-effective on-device model inference. Various post-optimization techniques are used to reduce the model size and precision bits. However, these techniques often result in a significant reduction in performance. To solve these issues, we propose a quantization-aware training (QAT) strategy for optimizing the CNNs to low-bit integers, resulting in faster inference and less memory utilization. We inject fake quantization modules into the original architecture, train the model in complete precision, and then convert the model to an 8-bit integer (INT8). The resultant QAT model performs all the computation of the convolution layers, activation layers, and batch-normalization in INT8. Our method reduces the size of ResNet50 and ResNet101 by a factor of 3.9x and improves the inference speed by more than 2x. We utilize the CIFAR-10 and CIFAR-100 datasets to test the performance of the models.

마이크로컨트롤러(Microcontroller Unit, MCU) 환경에서 연속학습(Continual Learning)은 새로운 태스크 학습 시 기존 지식의 망각(Catastrophic Forgetting) 문제와 제한된 메모리라는 이중 도전에 직면한다. 본 연구에서는 적응적 레이어 선택 기법과 저비트 양자화를 결합하여 메모리 효율적인 연속학습 방법을 제안한다. 실험 결과, 제안한 적응적 방법은 384KB 예산에서 고정 방법 대비 망각률을 15.3%p 감소시켰으며, SIEF-QPG 양자화는 기존 INT8 대비 손실 분산을 17.1% 개선하였다. 또한 동적 선택 모드는 정적 모드 대비 평균 3.8%p 높은 정확도를 달성하면서 수렴속도를 9.2% 향상시켰다. 본 연구는 자원 제약 엣지 디바이스에서의 지속적 모델 적응 가능성을 제시한다.

In microcontroller unit (MCU) environments, continual learning faces the dual challenges of catastrophic forgetting when learning new tasks and limited memory. This study proposes a memory-efficient continual learning method by combining an adaptive layer selection technique with low-bit quantization. Experimental results show that the proposed adaptive method reduced the forgetting rate by 15.3 percentage points compared to the fixed method within a 384KB budget. SIEF-QPG quantization improved loss variance by 17.1% compared to conventional INT8. Furthermore, the dynamic selection mode achieved an average accuracy 3.8 percentage points higher than the static mode while improving convergence speed by 9.2%. This research demonstrates the feasibility of continuous model adaptation on resource-constrained edge devices.

본 연구는 학교 현장에서 활용 가능한 한국어 기반 경량화 LLM(sLLM) 개발 및 적용 방안을 제시하고자 한다. 현재 상용 LLM 서비스는 개인정보 유출 위험, 서버 접속 환경 제한, 비용 문제 등으로 교육 현장에서 활용하기에 어려움이 있다. 이를 해결하기 위해 한국어 데이터셋을 활용한 파인튜닝과 양자화를 통해 학교에 보급된 스마트기 기에서 구동 가능한 경량화 LLM을 개발하고자 하였다. 연구 결과, Gemma-3-4b LLM을 KorQuAD 2.0데이터셋으 로 파인튜닝하고 Q5_K_M 형식으로 양자화한 LLM이 다른 LLM과 비교하여 한국어 이해력 측면에서 우수한 성 능을 보였다. 그리고 개발한 모델을 학교에 보급된 스마트기기에서 실행하여 프롬프트에 대한 결과를 응답과 초당 토큰 값을 기준으로 정리하여 교육적 활용 가능성을 확인하였다.

This study aims to suggest the development and application of Korean-based lightweight LLM (sLLM) that can be used in the school field. Currently, commercial LLM services are difficult to use in the educational field due to the risk of personal information leakage, server access environment restrictions, and cost problems. To solve this problem, we tried to develop a lightweight LLM that can be operated in smart devices distributed to schools through fine tuning and quantization using Korean datasets. As a result of the study, LLM, which was fine-tuned with the KorQuAD 2.0 dataset and quantized in the form of Q5_K_M, showed superior performance in terms of Korean understanding compared to other LLM. And the developed model was executed on smart devices distributed to schools, and the results of prompts were organized based on response and token values per second to confirm the possibility of educational use.

곱셈은 기본적인 산술 연산자로서 디지털 신호처리, 영상처리 등에서 널리 사용된다. 곱셈기는 면적과 전 력소모가 크므로 디지털시스템 구현 시 효율적인 곱셈기의 설계는 필수적이다. 고정길이 곱셈기는 일반 곱셈기 보 다 면적과 전력소모에서 우수하나, 양자화로 인한 최대오차에 때문에 사용이 제한된다. 본 논문에서는 고정길이 radix-8 Booth 곱셈기의 설계 방법을 제안하고, 제안한 곱셈기의 오차범위를 분석한 후 최대오차를 감소하기 위한 방법을 제안한다. 고정길이 곱셈기의 오차 성능 비교 시뮬레이션에서 제안한 방법의 오차 성능이 기존 방법보다 우 수하였다. 또한, 제안한 고정길이 radix-8 곱셈기를 Verilog HDL로 설계하고 인텔 FPGA로 합성하여 비교한 결 과, 절사후 고정길이 곱셈기 보다 하드웨어는 약 35%, 지연시간은 13%, 전력소모는 약 23% 적게 사용하였다.

Multiplication is a basic arithmetic operator and is widely used in digital signal processing, image processing, and so on. Since the multiplier has a large area and power consumption, the design of an efficient multiplier is essential in digital system implementation. A fixed-width multiplier is superior to a general multiplier in area and power consumption, but its use is limited due to the maximum error caused by quantization. In this paper, we propose a design method of a fixed-width radix-8 Booth multiplier and analyze the error bound of the proposed multiplier. Then, we propose a method to reduce the maximum error of the proposed multiplier. Through a comparative simulation of the error performance of the fixed-width multiplier, the error performance of the proposed method is superior to that of the existing methods. In addition, as a result of designing the proposed multipliers in Verilog HDL and synthesizing those in Intel FPGA, it can be seen that the proposed multipliers occupy about 35% less hardware, 13% less delay time, and 23% less power consumption compared to post truncated multipliers.

Due to the recent improvement in the computational power of hardware and the growth of data, a deep learning-based approach that optimizes parameters using massive data showed excellent performance. In computer vision, research using a convolutional neural network(CNN) is being actively conducted. However, it is challenging to apply to IoT devices due to the high computational complexity and massive memory usage required. In this paper, we propose a quantized CNN hardware for IoT devices that optimized memory usage and computation complexity. In addition, we present a quantization framework for the proposed hardware design. The presented framework includes floating-point training, quantization, fully integer arithmetic inference, and hardware design processes. As a result of implementing the quantized CNN on the Xilinx ZC702 evaluation board, power consumption and inference speed improved by 4.86× and 2.58×, respectively, compared to 32-bit floating-point hardware.

According to Unification Thought, Original laws, referred to as fundamental laws, may manifest themselves as the laws of the natural world. However, there can be found neither examples of originals laws nor any demonstration of such manifestation. It is my research interest to identify original laws and reveal their manifestation as the laws of the natural world. In this study, I have focused on quantum mechanics as the law of the natural world and traced back toward the source giving rise to such a phenomenon as a way of identifying an original law. In this process, numerous examples in the history of science as well as of civilization are considered in view of God-given creativities and application capabilities according to Unification Thought.

Data-Free Quantization (DFQ) enables model quantization without real data. Therefore, the accuracy of DFQ is largely affected by generated samples. However, we find that confidence distributions predicted from full-precision model over real and synthetic samples are significantly dissimilar, and this distribution discrepancy undermines quantization accuracy. To resolve this, we present an entropy regularization during synthetic data generation to make it resemble to real one much more. Furthermore, we propose Scaled Logit Alignment during quantization-aware training to bridge the representational gap between models. Our method achieves superior performance compared to the recent DFQ methods on ViT, CNN, and object detector architectures.

Testing the effectiveness of quantization matrices in (image) compression is difficult, because each image is unique and requires individualization techniques to optimize the compression ratio. However, there is usually no time to perform these kinds of techniques each time an image is compressed, and thus more general (but less effective) quantization matrices are often used. Although these matrices are standardized, there is still room for improvement. This paper proposes the usage of a popular machine learning technique, genetic algorithms, to actually perform this improvement. Although there might be some generalization issues, we believe it can be partially overcome if the right training set is chosen.

스마트폰 대중화의 영향으로 각종‘애플리케이션’이 개발되고 있으며 국악을 활용한 앱(app)도 출시되고 있으나 그수와 종류가 다양하지 않다. 특히 음악을 소재로 한 앱(app)의 경우 연주 실감도가 높아야 상품성을 확보할 수 있으며 이를 위해서는 실제 연주에 가깝게 작동되어야 하지만 실감연주에 적용할 국악 데이터는 미비한 실정이다. 따라서 본연구는 컴퓨터 음악(MIDI)으로 제작하는 국악 장구장단의 연주 실감도를 높이기 위해 실제 연주를 기반으로 타점과 음량에 적용할 수 있는 데이터를 제시하는 것이며, 이를 위해 실제 장구장단 연주의 타점(ms)과 음량(dB)을 악보기반 데이터와 비교분석하였다. 또한 실제 연주에 기반 한 타점의 위치를 정량화하고, 1박을 960으로 분해한 tick 데이터와 상대적 강약(Velocity)을 산출하여 모든 템포에 적용 가능하도록 하였다.

Thanks to the popularization of smartphone, a variety of applications have been developed including some of them handling “Guk-ak”. However, Guk-ak applications currently lack both diversity and usability. In general, music apps must enable realistic performances via sequencing close to real performances in order to increase merchantability. Guk-ak apps, however, failed to satisfy such requirement. Therefore, this research has its objective in proposing data on beating timing and velocity based on an actual performance to enhance performance liveliness of Guk-ak Jang-gu Jang-dan produced via MIDI.

정확한 인식률을 보이고 있는 상업적인 음성인식 시스템은 화자종속 고립데이터로부터 학습 모델을 사용한다. 그러 나 잡음 환경에서 데이터양에 따라 음성인식의 성능이 저하되는 문제점이 있다. 본 논문에서는 가우시안 분포에서 Maximum Log Likelihood를 이용한 벡터 양자화 기반 음성 인식 성능 향상을 제안한다. 제안하는 방법은 음성에 대한 특징 을 가지고 벡터 양자화와 Maximum Log Likelihood 음성 특징 추출 방법을 이용하여 유사 음성에 대한 음성 인식의 정확성 을 높이는 최적 학습 모델 구성 방법이다. 이를 위해 HMM을 기반으로 음성 특징을 추출하는 방법을 사용한다. 제안하는 방법을 사용하여 기존 시스템에서 생성되어 사용되는 음성 모델에 대한 부정확한 음성 모델에 대한 정확성을 향상시킬 수 있으므로 음성 인식에 강인한 모델을 구성할 수 있다. 제안하는 방법은 음성 인식 시스템에서 향상된 인식의 정확도를 보인다.

Commercialized speech recognition systems that have an accuracy recognition rates are used a learning model from a type of speaker dependent isolated data. However, it has a problem that shows a decrease in the speech recognition performance according to the quantity of data in noise environments. In this paper, we proposed the vector quantization based speech recognition performance improvement using maximum log likelihood in Gaussian distribution. The proposed method is the best learning model configuration method for increasing the accuracy of speech recognition for similar speech using the vector quantization and Maximum Log Likelihood with speech characteristic extraction method. It is used a method of extracting a speech feature based on the hidden markov model. It can improve the accuracy of inaccurate speech model for speech models been produced at the existing system with the use of the proposed system may constitute a robust model for speech recognition. The proposed method shows the improved recognition accuracy in a speech recognition system.

여러 분야에서 뛰어난 성능을 보이는 딥러닝은 뉴럴 네트워크의 은닉 계층을 늘려 깊은 네트 워크 구조를 형성할 수 있다. 이에 따라 복잡한 데이터를 쉽게 분류할 수 있으나, 가중치 수 증가로 인해 학습 연산량 및 메모리가 증가한다. 이처럼 많은 연산과 메모리가 필요한 딥러닝은 일반적으로 클라우드 상으로 학습하나, 클라우드는 사용자와의 통신상태가 원활해야 하며 서비스 비용에 대한 부 담감 및 개인정보 유출에 대한 위험성을 가진다. 이러한 문제를 해결하기 위해서 딥러닝을 임베디드 디바이스에 탑재해야 하며 온디바이스 탑재를 위한 네트워크 경량화가 필요하다. 경량화 기법으로는 프루닝 기법과 양자화 기법이 널리 사용되고 있으며 기존 양자화 기법은 비지도 학습 중 하나인 Kmeans를 사용하여 각 계층의 가중치들을 군집화하여 대푯값을 결정한다. K-means 기법에서 군집의 개수를 의미하는 K는 시행착오 방법을 통해 설정해야 하며 이에 따른 오버헤드가 발생한다. 또한, 기 존 양자화 기법은 네트워크 각 계층의 연결 관계를 무시한 채 독립적으로 양자화한다. 우리는 연산 오 버헤드 문제를 해결하기 위해 가중치들의 통계적 정보를 이용한 즉각적 가중치 양자화 기법을 제안하 고 네트워크 각 계층의 의존관계를 고려하여 순차 적으로 양자화함으로써 향상된 학습 성능을 보여준다.

Deep learning, which has excellent performance in various fields, can form deep network structure by increasing hidden layer of neural network. Thus, complex data can be easily classified, but the amount of learning computation and memory increases due to increase of the number of weights. Deep learning, which requires a lot of computation and memory, usually is trained on cloud environments. However, the cloud environment must be maintained in good communication state with user and have burden on service cost and risk of leakage of personal information. To solve this problem, deep learning needs to be learnt on the embedded device, and be light-weighting. Pruning and quantization methods are widely used for the light-weighting. Existing quantization methods usually use K-means, one of the unsupervised learning methods, in order to determine representative values of clustering weights. In the K-means method, the number of groups is set through trial-and-error method, which results in the computation overhead. To alleviate the overhead, we propose an instant weight quantization method using statistical information of weights Also, it can improve learning performance by quantizing sequentially with considering the dependency between layers in the network.

연합학습(Federated Learning)은 로컬데이터 기반의 학습 파라미터를 서버로 보내어 파라미터를 업데이트하는 방식의 분산 학습 방식으로써 로컬 데이터 유출을 방지하면서 모델을 공동 학습할 수 있는 구조를 제공한다. 연합학 습은 로컬 데이터의 유출은 방지하지만 파라미터 전송으로 인한 네트워크 부하와 파라미터에 대한 역복원 공격 (Inversion Attack)에 취약하다는 한계가 있다. 본 연구는 연합학습에서 파라미터 전송 효율성과 보안성을 향상시 키기 위해 확장 양자화 방식을 적용한 성능 분석을 다룬다. 기존의 gradient 클리핑, 차등 프라이버시, 양자화 기법 은 역복원 공격에 대한 보안성을 높이지만 전송량 절감 효과에는 제한적이다. 본 연구는 양자화 시 값의 범위를 확 장하여 양자화를 수행함으로써 유효 파라미터값을 줄이는 확장 양자화 기법을 적용하여 보았고, 다양한 확장 비율과 양자화 레벨에 따른 전송량 절감 및 역복원 공격 방어 성능을 실험적으로 분석했으며, 실험 결과, 확장 양자화는 약 간의 정확도 손실을 감수하는 대신 전송량을 효과적으로 줄이고 보안성을 향상시킬 수 있음을 확인하였다.

Federated Learning is distributed learning method in which locally trained parameters are sent to a server for parameter updates, enabling collaborative model training while preventing the leakage of local data. Although it protects local data from exposure, Federated Learning still suffers from limitations such as network overhead caused by parameter transmission and vulnerability to inversion attacks on the transmitted parameters. This study analyzes the performance of applying an extended quantization method to improve both parameter transmission efficiency and security in Federated Learning. While existing techniques such as gradient clipping, differential privacy, and standard quantization can enhance security against inversion attacks, their effectiveness in reducing transmission volume is limited. In this work, we applied an extended quantization technique that reduces the number of effective parameter values by expanding the quantization range, and we experimentally analyzed its impact on transmission reduction and defense against inversion attacks across various expansion ratios and quantization levels. Experiments show that extended quantization can effectively reduce transmission volume and improve security, with the cost of small loss in model accuracy.

Multi-view system is one of the auto-stereoscopic displays that allows users to watch 3D images without wearing special glasses. Generating contents for the multi-view display system is often accomplished using Depth-image-based rendering (DIBR) intermediate view generation. DIBR-based intermediate view generation method reconstructs a 3D scene with a color image and depth map information. Then, a virtual camera is put in various view positions to capture intermediate view images. In this research, we tried to improve the quality of the DIBR-based intermediate views by applying non-uniform quantization method for depth map information. This paper describes the intermediate view generation process using DIBR. Then, it discusses the evaluation of multi-view intermediate images generated using uniform and non-uniform depth quantization methods. Finally, the paper describes the future direction and how the muti-view systems can be utilized in computer games.

본 논문에서는 새로운 ROM 압축방식을 사용한 저전력 직접 디지털 주파수 합성기를 제안하고 낮은 클럭에서 동작하는 위상 누적기를 병렬로 연결하여 높은 주파수를 생성하는 위상-사인 변환기를 설계한다. ROM크기를 줄이기 위해 사인파를 양자화 할 때 일련의 차동 양자화 기술을 응용, 변형하여 양자화 ROM(Quantized ROM : Q-ROM과 차동 ROM(Differential ROM : D-ROM)을 사용하는 QD-ROM 압축방식을 제안한다. 이를 사용함으로써 67.5%의 ROM 사이즈를 감소시킬 수 있고 ROM의 크기를 줄여 전력 소모를 줄일 수 있다.

In this paper, a new method to reduce the size of ROM in the direct digital frequency synthesizer(DDFS) is proposed. And we design the phase-to-sine converter using the phase accumulator of parallel type for generating the high frequency. The new ROM compression method can reduce the ROM size by using the two ROM. The quantized value of sine is saved by the quantized-ROM(Q-ROM) and the differential ROM(D-ROM). So the total size of the ROM in the proposed DDFS is significantly reduced compared to the original ROM. The ROM compression ratio of 67.5% is achieved by this method. Also, the power consumption is decreased according to the ROM size reduction.

Unauthorized copying and distribution of digital audio has been greatly facilitated by the wide availability of low-cost personal computers, portable devices, network access, and audio recording and editing software. One possible solution for copyright protection is audio watermarking. In this paper, we propose an effective and robust audio watermarking algorithm that employs both the discrete wavelet transform and the discrete cosine transform. The algorithm involves, first, pre-processing of the binary watermark image and then embedding it into the original audio by quantization of coefficients. Experiments on audio recordings of many different music styles confirm the robustness of the algorithm against a wide range of Stirmark attacks such as noise addition, compression, and filtering, as well as other common attacks.

We proposed a new non-uniform sampling and quantization using variables low-pass filter with correlation. It focuses on the naturalness and intelligibility of speech synthesis applications and the compression and signal-to-noise ratio of speech transmission applications. However, it is well known that when conventional sampling methods are applied directly to speech signal, the required amount of data is comparable to or more than that of uniform sampling method. To overcome this problem, a new non-uniform methods is proposed, in which time domain coding is applied to two low-pass filters in lower bandwidth and the remain signals are compensated by the Gaussian white signal, which is used to get high quality speech by correlation of signal .

With spreading of digital contents, stealing other’s properties by unpermitted people have become easier. In order to protect the legitimate rights of owners, the digital watermark technologies are becoming more popular. In this paper, we propose an audio watermarking algorithm based on DCT coefficients and audio feature quantization. The proposed algorithm protects watermark information using several factors, such as filtering, audio feature extraction formula, quantization value, etc. Furthermore, in order to evaluate the performance of our proposed algorithm, we experiment our algorithm under the attacks such as low pass filtering, noise addition, and compression. From the experiment result, we can see that the proposed algorithm has high robustness.

딥 러닝의 발전으로 다양한 AI 기반의 응용이 많아지고, 그 모델의 규모도 매우 커지고 있다. 그러나 임베디드 기기와 같이 자원이 제한적인 환경에서는 모델의 적용이 어렵거나 전력 부족 등의 문제가 존재한다. 이를 해결하기 위해 서 클라우드 기술 또는 오프로딩 기술을 활용하거나, 모델의 매개변수 개수를 줄이거나 계산을 최적화하는 등의 경량화 방법이 제안되었다. 본 논문에서는 다양한 프레임워크들의 상호 교환 포맷으로 사용되고 있는 ONNX(개방형 신경망 교환 포맷) 포맷에 딥러닝 경량화 방법 중 학습된 모델의 양자화를 적용한다. 경량화 전 모델과의 신경망 구조와 추론 성능을 비교하고, 양자화를 위한 다양한 모듈 방식를 분석한다. 실험을 통해 ONNX의 양자화 결과, 정확도는 차이가 거의 없으며 기존 모델보다 매개변수 크기가 압축되었으며 추론 시간 또한 전보다 최적화되었음을 알 수 있었다.

Due to the development of deep learning and AI, the scale of the model has grown, and it has been integrated into other fields to blend into our lives. However, in environments with limited resources such as embedded devices, it is exist difficult to apply the model and problems such as power shortages. To solve this, lightweight methods such as clouding or offloading technologies, reducing the number of parameters in the model, or optimising calculations are proposed. In this paper, quantization of learned models is applied to ONNX models used in various framework interchange formats, neural network structure and inference performance are compared with existing models, and various module methods for quantization are analyzed. Experiments show that the size of weight parameter is compressed and the inference time is more optimized than before compared to the original model.

Cassandra database system is one of the universal databases. To achieve high performance, we should allocate memory space rationally according to actual demands. Otherwise, it will influence reading and writing performance. Actually, we always allocate memory space according to experience and repeated attempts which usually won’t give us the best answer. To solve this problem, firstly we analyze the reading and writing processing of the Cassandra database and find out the corresponding memory space which will influence system performance. Secondly, we build up a relationship model between system performance and memory allocation and name it as The Memory Model of Reading and Writing Performance. We have already applied the relationship model to real database servers to guide memory allocation and performance evaluation. Simulation results show that this memory model could well describe the quantization relationship of memory space and system performance.