Carsten Trinitis’s research while affiliated with Technical University of Munich and other places

... Trajectory matching-based methods [2] optimize the synthetic dataset by aligning the parameter update trajectories of models trained on the original and synthetic data. Building on this, Du et al. [8] reduced the error in the distilled dataset by minimizing the accumulated trajectory, while Liu et al. [22] proposed an automated approach to match the closest synthetic data trajectories to the original data trajectories, improving the performance of the distilled data. These advancements collectively illustrate the progressive efforts in the field of dataset distillation to enhance efficiency, scalability, and performance. ...

Reference:

Robust Dataset Distillation by Matching Adversarial Trajectories

... Researchers have looked into integrating user-space networking technologies like XDP (Express Data Path) and DPDK (Data Plane Development Kit) to improve the performance of DDS middleware in order to get beyond the real-time performance limits. Bode et al. [7] showed in their study that CycloneDDS, in conjunction with DPDK, greatly lowers latency and boosts throughput, resolving some of the real-time issues with conventional DDS implementations. CycloneDDS and DPDK operate together to reduce mean latency by up to 31%, which makes it an attractive option for high-performance, networked real-time systems like largescale robotic systems or high-frequency trading. ...

Reference:

FastDDS-Based Middleware System for Remote X-Ray Image Classification Using Raspberry Pi

... Using computational models to improve energy-(by pre-processing HSI images on-board satellites) and communication efficiency of CubeSats are gaining popularity, as shown by authors in [11], [12] respectively. Model compression techniques such as quantisation are used to reduce the computational complexity (and thus energy consumption) of neural network models for deployment in space applications [13], [14]. Quantised deeplearning accelerators on Intel Movidius Vision Processing Units (VPUs) have been integrated into onboard HSI imaging systems for detecting artefacts within the captured images, before transmitting them to ground stations [15]. ...

Reference:

An Energy-Efficient Artefact Detection Accelerator on FPGAs for Hyper-Spectral Satellite Imagery

... This enables the satellite to downlink, for instance, only valuable non-cloudy data, thus optimizing the usage of the often limited communication channel [13]. On-board cloud segmentation can also play a role in reducing decision-making latency, increasing satellite autonomy with more optimal operations [12,13,15,25,26,27,28]. Furthermore, from a broader perspective, the models first trained for cloud binary segmentation can be retrained to solve more complex problems instead, e.g., multi-class segmentation, where each image pixel is assigned to one of various categories. ...

Reference:

Semantic Segmentation in Satellite Hyperspectral Imagery by Deep Learning

... Therefore, key performance metrics include transmission latency, CPU usage, memory consumption, and message loss rate. This paper compares the middleware used in each platform: DDS, i.e., FastDDS [6], Fig. 9. Middleware communication performance results based on data size, frequency, and number of subscribers. The first row represents data size, the second row represents frequency, and the third row represents the number of subscribers. ...

Reference:

Open-Source Autonomous Driving Software Platforms: Comparison of Autoware and Apollo

... It provides a standard interface that smoothly integrates the various system components. We chose FastDDS in this paper because of its scalability and high performance [4]. Additionally, FastDDS offers reliable, low-latency data transfer, which is very important in real-time communication. ...

Reference:

FastDDS-Based Middleware System for Remote X-Ray Image Classification Using Raspberry Pi

... This ensures that raw data remains local while only distilled knowledge is shared. Observe that our proposal takes a different perspective from the existing Federated Distillation (FD) schemes [7,2,8]. With the objective of improving FL performance and privacy guarantees, FD distills knowledge from multiple local models independently and transfers only compact, distilled information to the server that trains a global model using this distilled data. ...

Reference:

Secure Federated Data Distillation

... In [28], the authors integrate GPU partitioning with power-aware scheduling to optimize resource allocation in CPU-GPU heterogeneous systems under power constraints. Leveraging MIG for fine-grained workload co-location, their approach employs scalability and interference models to improve efficiency and achieve near-optimal configurations across diverse workloads. ...

Reference:

A Multi-Objective Framework for Optimizing GPU-Enabled VM Placement in Cloud Data Centers with Multi-Instance GPU Technology

... This enables the satellite to downlink, for instance, only valuable non-cloudy data, thus optimizing the usage of the often limited communication channel [13]. On-board cloud segmentation can also play a role in reducing decision-making latency, increasing satellite autonomy with more optimal operations [12,13,15,25,26,27,28]. Furthermore, from a broader perspective, the models first trained for cloud binary segmentation can be retrained to solve more complex problems instead, e.g., multi-class segmentation, where each image pixel is assigned to one of various categories. ...

Reference:

Semantic Segmentation in Satellite Hyperspectral Imagery by Deep Learning

... 네트워크 기술의 발전 및 반도체 성능 향상에 따라 단위 시간당 처리되는 트랜잭션의 양도 크게 증가하는 추세이다 [1][2][3]. 특히 인공 지능을 활용한 데이터 모니터링 및 자동화된 데이터 처리가 가능해짐에 따라 이런 추세는 더욱 가속화될 것이 예상된다. 특히, 모바일 디바이스를 이용한 트랜잭션 처리나 지능화된 온라인 지불 처리 응용에서 이런 현상이 가속화되고 있다 [3] [4]. ...

Reference:

A Novel Update Scheme for High-Performance Transaction Processing Systems with SSD Storage