Henrikas Giedra
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Henrikas Giedra has a total of 11 co-authors. Here are the top co-authors by number of shared publications:
- Dalius Matuzevičius
Department of Electronic Systems, Vilnius Gediminas Technical University (VILNIUS TECH), 10105 Vilnius, Lithuania
3 shared publications - Gabriela Vdoviak
Department of Electronic Systems, Vilnius Gediminas Technical University, Saulėtekio Ave. 11, LT-10223 Vilnius, Lithuania
3 shared publications - Tomyslav Sledevič
Department of Electronic Systems, Vilnius Gediminas Technical University, Saulėtekio Ave. 11, LT-10223 Vilnius, Lithuania
3 shared publications - Artūras Serackis
Department of Electronic Systems, Vilnius Gediminas Technical University, Saulėtekio Ave. 11, LT-10223 Vilnius, Lithuania
2 shared publications - Ervinas Gisleris
Department of Electronic Systems, Vilnius Gediminas Technical University, Saulėtekio Ave. 11, LT-10223 Vilnius, Lithuania
1 shared publication - Evaldas Bružė
Security Research Laboratory, Mykolas Romeris University, LT-08303 Vilnius, Lithuania
1 shared publication
Latest Publications
- Open Access
- 3 June 2026
- Published by MDPI AG in Electronics
- Vol. 15 (11) , 2451
Benchmarking Large Language Model Inference on Limited-Resource Edge Systems
Large language models (LLMs) are increasingly considered for deployment on edge and limited-resource systems, where local inference can reduce latency, improve privacy, and decrease dependence on cloud infrastructure. While prior studies have evaluated either task accuracy or hardware efficiency in isolation, few benchmarks combine generation-based response-quality evaluation with real-device power measurements on a representative limited-resource platform. This study addresses that gap by benchmarking twelve compact and mid-scale open-weight LLMs (sub-1B to 8B parameters), evaluating generation-based accuracy on a desktop platform and measuring deployment efficiency—throughput, power consumption, and energy use—on an NVIDIA Jetson Orin Nano Super; the accuracy–efficiency trade-off is thus established at the model-configuration level. Unlike prior Jetson-based evaluations relying solely on internal telemetry, this work pairs generation-compatible lm-eval accuracy tasks with a dual power-measurement setup that combines internal tegrastats rail readings with external board-level input power measured using a digital multimeter and explicitly compares GPU-accelerated and CPU-only inference modes. GPU-accelerated inference provided a clear advantage, increasing median throughput from 7.12 to 18.13 tok/s and improving external-meter energy efficiency from 0.453 to 0.823 tok/J, despite higher mean input power. Sub-1B models offered the best throughput and energy efficiency, whereas 7–8B models achieved stronger accuracy at a substantially higher energy cost per generated token. These results demonstrate that edge LLM deployment requires multi-objective evaluation balancing accuracy, throughput, power consumption, and energy efficiency.
- 3 June 2026
- Published by MDPI AG in Electronics
- Vol. 15 (11) , 2451
Open AccessBenchmarking Large Language Model Inference on Limited-Resource Edge Systems
Large language models (LLMs) are increasingly considered for deployment on edge and limited-resource systems, where local inference can reduce latency, improve privacy, and decrease dependence on cloud infrastructure. While prior studies have evaluated either task accuracy or hardware efficiency in isolation, few benchmarks combine generation-based response-quality evaluation with real-device power measurements on a representative limited-resource platform. This study addresses that gap by benchmarking twelve compact and mid-scale open-weight LLMs (sub-1B to 8B parameters), evaluating generation-based accuracy on a desktop platform and measuring deployment efficiency—throughput, power consumption, and energy use—on an NVIDIA Jetson Orin Nano Super; the accuracy–efficiency trade-off is thus established at the model-configuration level. Unlike prior Jetson-based evaluations relying solely on internal telemetry, this work pairs generation-compatible lm-eval accuracy tasks with a dual power-measurement setup that combines internal tegrastats rail readings with external board-level input power measured using a digital multimeter and explicitly compares GPU-accelerated and CPU-only inference modes. GPU-accelerated inference provided a clear advantage, increasing median throughput from 7.12 to 18.13 tok/s and improving external-meter energy efficiency from 0.453 to 0.823 tok/J, despite higher mean input power. Sub-1B models offered the best throughput and energy efficiency, whereas 7–8B models achieved stronger accuracy at a substantially higher energy cost per generated token. These results demonstrate that edge LLM deployment requires multi-objective evaluation balancing accuracy, throughput, power consumption, and energy efficiency.
- Open Access
- 23 April 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (9) , 4154
Modeling Narrative Activation and Affective Feedback in Ideologically Structured Telegram Ecosystems
Digital social platforms have transformed political discourse into complex socio-technical environments characterized by rapid narrative diffusion, emotional amplification, and large-scale audience interaction. Understanding how sentiment and semantic alignment interact within such environments is important for analyzing polarization and patterns of audience response. This study examines narrative–audience interaction in Telegram political ecosystems using a combination of sentiment analysis, semantic similarity measures, and engagement metrics. Transformer-based language models are applied to quantify relationships between source posts and user-generated comments, enabling joint analysis of affective tone and topical alignment. The results reveal a consistent affective–semantic asymmetry: user responses tend to remain semantically aligned with source narratives while shifting toward more negative sentiment. This pattern indicates that disagreement is predominantly expressed through affective reframing rather than through divergence from the original topic. Further analysis shows systematic differences across ideological groups. Pro-government channels exhibit higher reach and more stable discourse alignment, while pro-opposition channels generate stronger engagement and more pronounced negative sentiment shifts. Neutral channels display intermediate characteristics. These findings demonstrate that online political discourse in Telegram is characterized by stable topical anchoring combined with systematic variation in emotional response.
- 23 April 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (9) , 4154
Open AccessModeling Narrative Activation and Affective Feedback in Ideologically Structured Telegram Ecosystems
Digital social platforms have transformed political discourse into complex socio-technical environments characterized by rapid narrative diffusion, emotional amplification, and large-scale audience interaction. Understanding how sentiment and semantic alignment interact within such environments is important for analyzing polarization and patterns of audience response. This study examines narrative–audience interaction in Telegram political ecosystems using a combination of sentiment analysis, semantic similarity measures, and engagement metrics. Transformer-based language models are applied to quantify relationships between source posts and user-generated comments, enabling joint analysis of affective tone and topical alignment. The results reveal a consistent affective–semantic asymmetry: user responses tend to remain semantically aligned with source narratives while shifting toward more negative sentiment. This pattern indicates that disagreement is predominantly expressed through affective reframing rather than through divergence from the original topic. Further analysis shows systematic differences across ideological groups. Pro-government channels exhibit higher reach and more stable discourse alignment, while pro-opposition channels generate stronger engagement and more pronounced negative sentiment shifts. Neutral channels display intermediate characteristics. These findings demonstrate that online political discourse in Telegram is characterized by stable topical anchoring combined with systematic variation in emotional response.
- Open Access
- 11 July 2025
- Published by MDPI AG in Electronics
- Vol. 14 (14) , 2796
Deploying Optimized Deep Vision Models for Eyeglasses Detection on Low-Power Platforms
This research addresses the optimization and deployment of convolutional neural networks for eyeglasses detection on low-power edge devices. Multiple convolutional neural network architectures were trained and evaluated using the FFHQ dataset, which contains annotated eyeglasses in the context of faces with diverse facial features and eyewear styles. Several post-training quantization techniques, including Float16, dynamic range, and full integer quantization, were applied to reduce model size and computational demand while preserving detection accuracy. The impact of model architecture and quantization methods on detection accuracy and inference latency was systematically evaluated. The optimized models were deployed and benchmarked on Raspberry Pi 5 and NVIDIA Jetson Orin Nano platforms. Experimental results show that full integer quantization reduces model size by up to 75% while maintaining competitive detection accuracy. Among the evaluated models, MobileNet architectures achieved the most favorable balance between inference speed and accuracy, demonstrating their suitability for real-time eyeglasses detection in resource-constrained environments. These findings enable efficient on-device eyeglasses detection, supporting applications such as virtual try-ons and IoT-based facial analysis systems.
- 11 July 2025
- Published by MDPI AG in Electronics
- Vol. 14 (14) , 2796
Open AccessDeploying Optimized Deep Vision Models for Eyeglasses Detection on Low-Power Platforms
This research addresses the optimization and deployment of convolutional neural networks for eyeglasses detection on low-power edge devices. Multiple convolutional neural network architectures were trained and evaluated using the FFHQ dataset, which contains annotated eyeglasses in the context of faces with diverse facial features and eyewear styles. Several post-training quantization techniques, including Float16, dynamic range, and full integer quantization, were applied to reduce model size and computational demand while preserving detection accuracy. The impact of model architecture and quantization methods on detection accuracy and inference latency was systematically evaluated. The optimized models were deployed and benchmarked on Raspberry Pi 5 and NVIDIA Jetson Orin Nano platforms. Experimental results show that full integer quantization reduces model size by up to 75% while maintaining competitive detection accuracy. Among the evaluated models, MobileNet architectures achieved the most favorable balance between inference speed and accuracy, demonstrating their suitability for real-time eyeglasses detection in resource-constrained environments. These findings enable efficient on-device eyeglasses detection, supporting applications such as virtual try-ons and IoT-based facial analysis systems.
- Open Access
- 30 May 2024
- Published by Vilnius Gediminas Technical University in New Trends in Computer Sciences
- Vol. 2 (1) , 31-45
LEVERAGING GENERATIVE ADVERSARIAL NETWORKS TO IMPROVE TRAINING IMAGE DATASET
Convolutional neural networks (CNNs) are powerful models of deep learning that are widely used in computer vision classification tasks. The purpose of this study is to investigate the impact of datasets on CNN performance, employing original datasets and expanded datasets with synthetically generated images. The Generative Adversarial Network (GAN) is an unsupervised deep learning method used for synthetic data generation and can address the limitations of image augmentations. In this study, a new GAN architecture is used to synthesize high-resolution images when dealing with limited training data. The StyleGAN2-ADA model is specifically designed to generate high-quality images using limited datasets. Adaptive Discriminator Augmentation (ADA) dynamically adjusts data augmentation, enhancing discriminator efficiency and stability. The findings indicate a reduction in the likelihood of overfitting, enhancement in network generalization, mitigation of class imbalance concerns, and a concurrent increase in the accuracy and stability of network classification.
- 30 May 2024
- Published by Vilnius Gediminas Technical University in New Trends in Computer Sciences
- Vol. 2 (1) , 31-45
Open AccessLEVERAGING GENERATIVE ADVERSARIAL NETWORKS TO IMPROVE TRAINING IMAGE DATASET
Convolutional neural networks (CNNs) are powerful models of deep learning that are widely used in computer vision classification tasks. The purpose of this study is to investigate the impact of datasets on CNN performance, employing original datasets and expanded datasets with synthetically generated images. The Generative Adversarial Network (GAN) is an unsupervised deep learning method used for synthetic data generation and can address the limitations of image augmentations. In this study, a new GAN architecture is used to synthesize high-resolution images when dealing with limited training data. The StyleGAN2-ADA model is specifically designed to generate high-quality images using limited datasets. Adaptive Discriminator Augmentation (ADA) dynamically adjusts data augmentation, enhancing discriminator efficiency and stability. The findings indicate a reduction in the likelihood of overfitting, enhancement in network generalization, mitigation of class imbalance concerns, and a concurrent increase in the accuracy and stability of network classification.
- Open Access
- 30 May 2024
- Published by Vilnius Gediminas Technical University in New Trends in Computer Sciences
- Vol. 2 (1) , 1-18
REVIEW AND EXPERIMENTAL COMPARISON OF GENERATIVE ADVERSARIAL NETWORKS FOR SYNTHETIC IMAGE GENERATION
The application of machine learning algorithms has become widespread particularly in fields such as medicine, business, and commerce. However, achieving accurate classification results with these algorithms often relies on large-scale training datasets, making data collection a lengthy and complex process. This paper reviews the current utilization of generative adversarial network (GAN) architectures and discusses recent scientific research on their practical applications. The study emphasizes the significance of addressing data scarcity in the process of training the machine learning algorithms and highlights the potential of advanced GAN architectures, in particular StyleGAN2-ADA, to mitigate this challenge. The findings contribute to ongoing efforts aimed at enhancing the efficiency and applicability of artificial intelligence across diverse domains by presenting a viable solution to the constraint of limited training data for image classification tasks.
- 30 May 2024
- Published by Vilnius Gediminas Technical University in New Trends in Computer Sciences
- Vol. 2 (1) , 1-18
Open AccessREVIEW AND EXPERIMENTAL COMPARISON OF GENERATIVE ADVERSARIAL NETWORKS FOR SYNTHETIC IMAGE GENERATION
The application of machine learning algorithms has become widespread particularly in fields such as medicine, business, and commerce. However, achieving accurate classification results with these algorithms often relies on large-scale training datasets, making data collection a lengthy and complex process. This paper reviews the current utilization of generative adversarial network (GAN) architectures and discusses recent scientific research on their practical applications. The study emphasizes the significance of addressing data scarcity in the process of training the machine learning algorithms and highlights the potential of advanced GAN architectures, in particular StyleGAN2-ADA, to mitigate this challenge. The findings contribute to ongoing efforts aimed at enhancing the efficiency and applicability of artificial intelligence across diverse domains by presenting a viable solution to the constraint of limited training data for image classification tasks.