Cost adaptive scheduling schemes for object identification with DNN acceleration and video streaming in HPC environment

Abstract

With the development of sensing technology, object recognition and data streaming technologies are being used for visual data analysis and transmission in various situations such as satellite images or CCTV video. Recently, research on automation of object recognition based on DNN for input data streams has been actively conducted. In particular, in the case of object recognition that requires high accuracy and reliability, an attempt to use Explainable DNN is drawing attention. Through this, Explainable DNN presents a human-model interactive relearning framework in which a visual interpretation of the input data stream is automatically generated and the reader retrains the DNN through feedback. However, the interactive relearning framework still has problems that degrade the data analysis performance, so a solution is required.

In general, interactive relearning frameworks have the high computing complexity of Explainable DNNs. Because object recognition requires extremely high accuracy and reliability, higher analysis accuracy can be obtained through a large-scale DNN. However, large-scale DNNs containing a large number of network layers and channels have high computing complexity and high memory demands, resulting in long latency. In addition, data to which object recognition is basically applied has a high transmission delay and workload as a large stream such as a continuous ultra high definition image or high definition video. The increased computational complexity due to this transmission delay and high workload degrades the processing performance of the interactive relearning framework and slows the convergence of model accuracy.

In this paper, to solve this problem, we study a scheduling technique for accelerating the Explainable DNN-based interactive relearning framework in HPC environment, and propose a new scheduling technique that considers the complexity of Explainable DNN in FPGA-GPU environment. Existing scheduling techniques do not take into account the computational complexity of DNN processing, and performance is degraded due to wasted use of accelerators and unnecessary delays. To solve this problem, we newly define a time and energy model when processing Explainable DNN in an FPGA-GPU environment, and propose a new scheduling scheme to guarantee latency and minimize energy cost.

Next, we propose a configuration adaptation technique that realizes cost-effective DNN-based object recognition acceleration by adaptively adjusting configuration values ​​(e.g. resolution, frame rate) for large-scale input data streams of the interactive relearning framework. This proposal guarantees the optimal performance of DNN processing speed/resource usage and accuracy for the frame rate and resolution of large-scale input data. In particular, we apply a lightweight online profiling method that takes advantage of the dynamic characteristics of an object that solves the existing high-cost profiling problem. This is a technique that improves the processing speed and dramatically improves the use of accelerator resources while ensuring the same accuracy as the existing DNN processing for the given input data.

In addition, we propose an adaptive unlabeled data selection method based on active learning and semi-supervised learning to reduce the labeling cost of the interactive relearning framework. Existing active learning applies only a small number of labeled data to learning rather than a large number of unlabeled data, causing an overfitting problem. This makes it difficult to converge model accuracy above a certain level and causes a bottleneck in satellite image reading. In this paper, we propose a new method to increase the amount of effective training data and accelerate model accuracy convergence by applying trust-based adaptive data selection and quasi-supervised learning for unlabeled data.

Finally, we propose a chunk caching method based on a short-term time-varying (STV) request model for accelerating real-time transmission delay for large-scale input data streams of an interactive relearning framework. This proposal maximizes cache hits by efficiently utilizing the limited storage capacity of the framework and caching chunks of the data stream to be requested with high probability in advance. Since the real-time data stream follows the STV characteristics, the probability model of the existing caching technique based on the chunk request recording does not guarantee cache hit performance. The newly proposed scheme increases cache hits for real-time data streams and lowers transmission delay.

In conclusion, this dissertation newly proposed cost-adaptive scheduling techniques for real-time transmission delay acceleration of large-scale input data streams in an interactive re-learning framework, Explainable DNN-based analysis, and learning acceleration. It was verified and showed excellence through experiments.