SERMAS: Magnetic Resonance Image Processing

 
The new technological advances in Magnetic Resonance Imaging (MRI) allow for the collection of vast amounts of imaging data in shorter acquisition times. However, raw imaging data rarely provides the metrics of interest for research studies or even for clinical use. This is especially true for Diffusion-Weighted Imaging (DWI) of the brain. In this modality, large numbers of volumetric images must be processed to provide information about the structural connectivity of the brain and metrics associated with white-matter integrity. This image processing is computationally expensive, taking several hours for only one subject in a standard computer. However, most of the processing steps are highly parallelisable, not only per-subject but also in a per-volumetric-pixel fashion. Therefore, this kind of processing will significantly benefit from HPC and Exascale infrastructures, making it possible to decrease the processing times dramatically. This will open the possibility for more efficient and productive research in neuroimaging, enabling the use of more complex processing algorithms currently limited by computational power. Moreover, faster processing of MRI images will lead to a more rapid diagnosis of neural disorders.
 

UNICAL: Urban computing

 

Urban computing is the process of acquisition, integration, and analysis of big and heterogeneous urban data to tackle the major issues that cities face today, including air pollution, energy consumption, traffic flows, human mobility, environmental preservation, commercial activities and savings in public spending.

Data gathered from social media, such as posts from Twitter and Facebook or photos from Instagram and Flickr, are frequently geotagged. Such data can be analyzed for extracting valuable information about user mobility, including the most frequent sets of places visited by users and the most common trajectories followed by users.

The analysis of user trajectories through RoIs (Regions of interest) is highly valuable in many scenarios:

  • tourism agencies and municipalities can discover the most visited touristic places and the time of year when such places are visited;
  • transport operators can discover the places and routes where is it more likely to serve passengers and crowed areas where more transport facilities need to be allocated.
  • business analysts can analyze the flow of the users – including a departure and arrival points, path, transportations, waiting time – to suggest the best place where to open a business activity.

 

PNSC: Deep Learning

Cognitive computing is perceived as one of the main challenges in the current as well as future systems, from so-called edge computing on embedded platforms to high performance computing in the largest data centers. Among the possible approaches used to make a computer system understand raw data, deep learning is the one that is described as blooming. Deep learning attracted world-wide attention of both research and commercial communities which resulted in development of multiple frameworks that help researchers focus on the design of neural networks rather than on complex math and low-level execution performance. The most popular AI frameworks include: Caffe, TensorFlow, Torch , Theano and CNTK, among others. They are often used for image, voice, video and other raw data classification, detection and segmentation. There are three factors that have enabled the advent of deep learning: 1) algorithms and new types of networks, 2) immense processing power and 3) availability of big data. The last factor refers to the fact that deep learning would not exist if it wasn’t for large volumes of training data. The idea is that instead of having a team of researchers extracting custom features from data, we provide a massive amount of data and let the software automatically learn from it. This is where the ASPIDE project may come into action: the optimal placement of data with respect to the processing power is crucial in order to achieve high and sustainable learning performance at Exascale.
 
PSNC is using deep learning in a few scenarios related to image
recognition and understanding of time-series data.
 

Image Recognition

The objective of image recognition use case is to create a model able to recognise Varroa mites in images made by beekeepers. This model is going to be a part of a system supporting beekeepers in choosing type of treatment.

Time series analysis

Time series use case focus on analysis of data from electric vehicles and factories. Each of them can produce a massive amount of data from number of parameters. Such data, can contain information about potential issues and untypical situations, but due to high amount, they cannot be analysed manually by experts. Thus, unsupervised anomaly detection approaches must be used.

Computational Fluid Dynamics

PSNC also develops a model, which should support or replace computationally demanding CFD simulations with Deep Learning approaches. Such model could be use for faster prototyping.

 

 

 

 

INTEGRIS: Big Data Analytics

Nowadays Big Data applications are disrupting the market of the data processing industry in multiple areas/sectors, including Broadcasting, Social Media and Telecom. Most of the big data systems currently relies on the Hadoop ecosystem, which empowered the development of numerous powerful paradigms and tools such as HDFS, MapReduce, Spark, etc. Inevitably, those tools still have lacks and limitations, mainly due to the increasing amount of data involved and its distributed nature, such as frequent low performance scenarios, issue with small files, little to no interaction processing.

The proposed use case is based on the analysis of data from Telco sector. The system gathers data from the national cellular network, consisting of more than 120.000 cells.  Starting from a data set on HDFS, containing information about mobile users’ distribution on the Italian territory, the core of the application allows quantitative analysis by aggregation and filtering functions mainly following a Map and Reduce model. The Hadoop tools are used in the use case, but the limitations abovementioned often occur. The aim of the ASPIDE project is to overcome these boundaries with the definition of new tools, programming paradigms, and methodologies.