Academic Computer Centre Cyfronet AGH
ACK Cyfronet AGH
Serwis używa plików cookies zgodnie z polityką prywatności. Pozostając w serwisie akceptują Państwo te warunki.

Laboratory of Acceleration of Computing and Artificial Intelligence

Our mission

Thanks to our expertise in the field of the Artificial Intelligence (AI) algorithms, and our knowledge of the modern computational methods, we can support scientific community in their AI-based research. Our knowledge and experience allows us to implement machine learning algorithms and dedicated to neural networks effectively using the AI-dedicated partition of Prometheus’ supercomputer available at ACC Cyfronet AGH. The AI partition of Prometheus was built based on the four efficient computing servers; each equipped with eight nVidia Volta V100 GPGPU cards. The total computing power of the partition is over 4 PetaFlops, which is over four quadrillions (4x1015) of AI-dedicated computations per second.

Domain specializations

Selected applications of Artificial Intelligence in the areas of natural language processing (NLP), image processing and time series analysis for various research problems are presented below.

Natural language processing

The Laboratory is proud of many years of experience in natural language processing. As an example, we have created and developed the tool that allows the users to search, compare and classify text documents. The result of this work is the web service called Scholar that is available at ACC Cyfronet AGH on the PLGrid platform. One of the important research problems is the analysis of impact of the methods for reducing the accuracy of textual data representation on the effectiveness of the NLP algorithms. We have managed to develop alteration of the methods that allow for a 10-fold reduction in computing energy consumption, if compared to the original implementation, with no significant loss of accuracy.

The emergence of neural networks-based solutions has revolutionized the NLP field. We research on the compression and hardware implementation of the sentiment assessment network, which showed that it is possible to reduce the accuracy of the network coefficients precision to 8 or even 4 bits while maintaining the network efficiency almost unchanged. Additionally, we examine the area of semi-supervised learning, where the amount of available tagged data is very limited and the output categories changed during the operation of the system. Our research has shown that it is possible to develop a solution with an accuracy of up to 98.9%.

Image processing

In the field of image processing using neural networks, the Laboratory's work focuses on the recognition and detection of objects for the needs of medical applications. During the tests, we developed a system for the classification of neoplastic changes in the samples from the cytological examination of the animal tissue. The proposed system achieved an efficiency of approximately 96% for the three selected types of cancer. We used deep network models based on Resnet-50 and Resnet-152 nets. Also, to allow for the selection of the perception area in classification operations, we have developed a special training scheme that is based on genetic algorithms. One of the latest developments in the area is the Yolo3-based detection, which achieves a very good value 0.86 of mAP for the specially prepared images that are made with the use of many low-quality cytological preparations. The system that is now developed is to be ultimately used in the daily work of veterinary clinics.

Time series analysis

The works of the Laboratory also concern the modelling of time series for the detection of unusual situations. Detailed work includes practical applications such as anomaly prediction to avoid catastrophic damage to the magnets and other associated devices of the Large Hadron Collider (LHC) at CERN. By means of GRU and LSTM neural networks and dedicated post-processing, we have developed a system for quench detection in superconducting magnets, which additionally enabled the classification of detected anomalies. Moreover, to allow the user for very low response latency, the developed algorithm was implemented on the computing platform that is based on the Xilinx Zynq UltraScale+ MPSoC 285 FPGA architecture.

In the field of medicine, we address the problem of prediction of fainting of hospitalized patients, who are confined to a hospital bed for a long time. Thus, the analyzes concern the well-known problem of the loss of the leg muscles support for the cardiovascular system. This research is carried out with the cooperation of the Medical University of Graz.

Contact: Paweł Russek, p.russek [at] cyfronet.pl