Перегляд за Автор "Kysil , T."

Зараз показуємо 1 - 2 з 2
  • Ескіз
    Документ
    Method for improving the performance of convolutional neural networks using an accelerator
    (Хмельницький національний університет, 2024) Isaiev, T.; Kysil , T.
    The effectiveness of convolutional neural networks (CNNs) has been demonstrated across various fields, including computer vision, natural language processing, medical imaging, and autonomous systems. However, achieving high performance in CNNs is not only a matter of model design but also of optimizing the training and inference processes. Using accelerators like the Google Coral TPU provides significant improvements in both computational efficiency and overall model performance. This paper focuses on the integration of the Coral TPU to enhance CNN performance by speeding up computations, reducing latency, and enabling real-time deployment. Training deep learning models, particularly CNNs, is computationally intensive. Traditional CPUs or GPUs can take hours or even days to train large networks on complex data. The accelerator offloads these intensive tasks, allowing the host machine to focus on other operations and making training more efficient. This enables researchers to experiment with multiple architectures and hyperparameters within shorter cycles, thereby improving the model's accuracy and robustness. CNNs are widely deployed in edge computing scenarios where real-time predictions are critical, such as in robotics, autonomous vehicles, and smart surveillance systems.Unlike traditional cloud-based solutions, where models are executed remotely and suffer from network delays, the Coral TPU ensures low-latency predictions directly on the device, making it ideal for timesensitive applications. Another key advantage of using accelerators like Coral TPU is the ability to efficiently handle optimized and lightweight models. These optimized models are well-suited for the Coral TPU’s architecture, allowing developers to deploy high-performing networks even on resource-constrained devices. The TPU’s ability to handle quantized models with minimal loss in accuracy further enhances the CNN’s practical usability across various domains. The Coral TPU is designed to minimize power consumption, making it an ideal solution for battery-powered or energyconstrained devices. This energy efficiency ensures that CNNs can run continuously on devices like drones, IoT sensors, or mobile platforms without exhausting their power supply. Additionally, the scalability of the TPU makes it easy to deploy multiple accelerators in parallel, further improving throughput for applications that require processing high volumes of data, such as realtime video analysis. The Coral TPU also facilitates on-device learning, where models can be incrementally updated based on new data without requiring a full retraining session. This feature is particularly useful in dynamic environments, such as autonomous vehicles or security systems, where the model needs to adapt quickly to new conditions. With the TPU handling the computational workload, CNNs can be fine-tuned on the device, ensuring they remain accurate and responsive over time.
  • Ескіз
    Документ
    Method of creating custom dataset to train convolutional neural network
    (Хмельницький національний університет, 2024) Isaiev, T.; Kysil , T.
    The task of creating and developing custom datasets for training convolutional neural networks (CNNs) is essential due to the increasing adoption of deep learning across industries. CNNs have become fundamental tools for various applications, including computer vision, natural language processing, medical imaging, and autonomous systems. However, the success of a CNN depends heavily on the quality and relevance of the data it is trained on. The datasets used to train these models must be diverse, representative of the task at hand, and of sufficient quality to capture the underlying patterns that the CNN needs to learn. Thus, building custom datasets that align with the specific objectives of a neural network plays a critical role in enhancing the performance and generalization capability of the trained model. This paper focuses on developing a method and subsystem for generating high-quality custom datasets tailored to CNNs. The aim is to provide a framework that automates and streamlines the processes involved in data collection, preprocessing, augmentation, annotation, and validation. Moreover, the method integrates tools that allow the dataset to evolve over time, incorporating new data to adapt to changing requirements or environments, making the system flexible and scalable. The process of creating a dataset begins with the acquisition of raw data. The data can come from various sources such as images from cameras, videos, sensor feeds, open data repositories, or proprietary datasets. A key consideration during data collection is ensuring that the samples cover the full range of conditions or classes the CNN will encounter in production. For example, in an object recognition task, it is essential to collect images from diverse environments, lighting conditions, and angles to train the model effectively. Ensuring variability in the dataset increases the model's ability to generalize, reducing the risk of poor performance on unseen data. Data augmentation is a critical step in building a robust dataset, particularly when the size of the dataset is limited. Augmentation techniques introduce variability into the dataset by artificially modifying the existing samples, thereby simulating a wider range of conditions. This helps the CNN generalize better and prevents overfitting. In essence, it allows the model to experience different perspectives and distortions of the same data, strengthening its adaptability to real-world scenarios. Annotation involves labeling the data samples with the correct class or category information. Depending on the task, annotations may include bounding boxes for object detection, segmentation masks for semantic segmentation, or class labels for classification tasks. The importance of well-annotated data cannot be overstated, as CNNs rely on this labeled information to understand the relationships between input data and the desired output predictions.