연합학습은 학습을 개별 클라이언트에 분산시켜 학습연산의 오버헤드는 분산되지만, 중앙 서버에 모델 파라미터를 전송하는 오버헤드가 증가한다. 본 연구에서는 연합학습에서 클라이언트와 중앙서버간 전송되는 파라미터의 양을 줄이기 위해서 오려낸(Clipped) 양자화를 파라미터에 적용하여 본다. 오려낸 양자화(Clipped Quantization)는 연합학습 시스템에서 클라이언트가 매 라운드 시 업데이트된 파라미터를 양자화하여 중앙서버에 보낼 때, 양자화 값의 분포를 파악하여 양자화된 파라미터의 양 끝부분을 일정 비율로 클리핑한 후 이 값을 압축하여 보내는 방식이다. 이러한 오리기 방법은 양자화된 값 중에서 유효한 값을 가지는 범위가 줄어듦으로써 압축률을 높일 수 있으므로 전송량을 줄일 수 있다. 실험결과, 딥러닝 모델에서 2~5%의 정확도 감소대비 30~40% 에폭 당 데이터 전송량이 줄어듦을 확인할 수 있었다.

Recently, IoT applications using Deep Neural Network (DNN) to embedded edge devices are increasing. Generally, in the case of DNN applications in the IoT system, training is mainly performed in the server and inference operation is performed on the edge device. The embedded edge devices still take a lot of loads in inference operations due to low computing resources, so proper customization of DNN with architectural exploration is required. However, there are few integrated frameworks to facilitate exploration and customization of various DNN models and their operations in embedded edge devices. In this paper, we propose an integrated framework that can explore and customize DNN inference operations of DNN models on embedded edge devices. The framework consists of the GUI interface part, the inference engine part, and the hardware Deep Learning Accelerator (DLA) Virtual Platform (VP) part. Specifically it focuses on Convolutional Neural Network (CNN), and provides integrated interoperability for convolutional neural network models and neural network customization techniques such as quantization and cross-inference functions. In addition, performance estimation is possible by providing hardware DLA VP for embedded edge devices. Those features are provided as web-based GUI interfaces, so users can easily utilize them.

Federated Learning(FL) has become an emerging method that trains private data by distributed learning of shared parameter, however, it has high communication overhead and is still exposed to attack on the model parameters. To minimize the communication overhead of federated learning while preserving its accuracy and security, we considered a combination of techniques with gradient quantization, clipping, and Huffman coding. Our system produces reduced parameters through quantization and clipping, then encodes the parameters through Huffman coding, which further increases the compression ratio as well as security of the parameters to be transmitted. Preliminary results identify that the scheme can significantly reduce the amount of transferring while preserving accuracy and security.

Recently, research on lightweighting of Deep Neural Network(DNN) models suitable for embedded IoT devices and edge computing is rapidly increasing. One of the important points of lightweighting DNN is model parameter optimization. Parameter precision increases model accuracy, but also increases memory and processing overhead. In this paper, we developed a parameter bitwise adjustment framework of DNN models to analyze the correlation between model’s parameter precision and performance. The framework can set each bit of parameter value to 0, 1, or random in a specific range for DNN models. With the framework we measured the accuracy of the model while modifying the parameter values to 0 by step by step from LSB, for three representative CNN models. As a result, it was identified that the accuracy robustness for the range of valid parameters is different each other depending on the model.

The Log-Structured Merge Tree (LSM Tree) is widely employed in key-value stores, ensuring efficient database read performance at the expense of increased write stall. While this stall enhances read performance, it notably degrades write efficiency. This paper examines the write stall phenomenon in LSM Tree-based key-value stores and proposes a solution: Key Space Partitioned RocksDB. This architecture comprises a MemTable backed by Storage DRAM, the Key Space Partitioned MemTable, and the Key Space Partitioned LSM Tree. Key Space Partitioned RocksDB demonstrates a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.7\times$</tex> enhancement in YCSB-A throughput compared to conventional RocksDB, a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3.5\times$</tex> reduction in average GET(key) latency, and a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.2\times$</tex> decrease in average PUT (key, value) latency.

최근 딥러닝 기술은 고정밀도 모델로 인해서 정확도가 높아지고 있으나, 복잡한 네트워크의 깊이와 파라미터의 수 증가로 인해서 연산에 필요한 자원 소모 및 프로세싱 시간이 증가한다. 이러한 네트워크 복잡도는 IoT 시스템 및 모바일 시스템과 같은 임베디드 시스템에서 딥러닝 연산을 수행하는 오버헤드가 된다. 따라서, 양자화와 같은 경량 시스템에 맞는 딥러닝 모델을 위한 경량화 연구가 많이 이루어져 왔다. 양자화를 적용하면 정확도 손실이 많이 발생하기 때문에 파라미터를 혼합해서 사용하는 Mixed Precision 적용이 대안이 되고있다. 본 논문에서는 임베디드 시스템에서 Mixed Precision 기반의 CNN 딥러닝 모델의 추론 연산 방식을 구현하였으며, Mixed Precision 적용에 대한 성능 분석을 하였다. 실험 결과, 연산 복잡도가 높은 몇몇 계층에 양자화를 적용하면 정확도의 손실을 줄이면서 추론 시간이 14%~20%가량 줄어듦을 확인할 수 있었다.

&lt;div class="section abstract"&gt;&lt;div class="htmlview paragraph"&gt;NVH analysis based on numerical simulations before actual test vehicle is available becomes common process in the automotive industry. Furthermore, the latest work scope is extending even to conceptual study in the very early design stage, beyond traditional numerical simulations simply using 3-D CAD data. In case when reasonable information is provided at this very early vehicle development stage, a better decision on the design concept would be possible, and subsequent design process can be carried out in more efficient manner. The core of this trend is that it allows us to predict vehicle performance at the conceptual design stage without 3-D CAD data, and then, with this prediction, to suggest meaningful design directions for next stage. From this point of view, FRF-Based Substructuring (FBS) methodology has potential to be used as an appropriate tool for this purpose. Because running FBS technique takes relatively very short computation time for an individual model, large amount of computations for various combinations are made possible, which are virtually impossible to implement with the conventional finite element analysis. In this research, a conceptual analysis in the early design stage on a small-sized passenger vehicle has been conducted using FBS method. After converting the Finite Element (FE) full vehicle into the FBS models of each major system and establishing the FBS model database with about three hundreds of combinations, it is examined to select the optimal combinations for improving the road-noise. Based on these results, the design guide lines on each system for the improvement of road-noise performance as well as cost reduction have been proposed before detail 3-D design data is released. In addition, it has been shown that the approach based on FBS database is also very useful in the target setting regarding NVH related performance on each individual system.&lt;/div&gt;&lt;/div&gt;

연합학습은 학습을 개별 클라이언트에 분산시켜 학습연산의 오버헤드는 분산되지만, 중앙 서버에 모델 파라미터를 전송하는 오버헤드가 증가한다. 본 연구에서는 연합학습에서 클라이언트와 중앙서버간 전송되는 파라미터의 양을 줄이기 위해서 오려낸(Clipped) 양자화를 파라미터에 적용하여 본다. 오려낸 양자화(Clipped Quantization)는 연합학습 시스템에서 클라이언트가 매 라운드 시 업데이트된 파라미터를 양자화하여 중앙서버에 보낼 때, 양자화 값의 분포를 파악하여 양자화된 파라미터의 양 끝부분을 일정 비율로 클리핑한 후 이 값을 압축하여 보내는 방식이다. 이러한 오리기 방법은 양자화된 값 중에서 유효한 값을 가지는 범위가 줄어듦으로써 압축률을 높일 수 있으므로 전송량을 줄일 수 있다. 실험결과, 딥러닝 모델에서 2~5%의 정확도 감소대비 30~40% 에폭 당 데이터 전송량이 줄어듦을 확인할 수 있었다.