Recent years have brought a large change in the way that computer vision is done, with the direction shifting from specialized feature description and rather simple algorithms to deep learning and its flavors like deep residual networks and convolutional neural nets. We present the implementation details of several methods in the framework of classifying traffic signs, comparing the results and formulating some insights as to what a good approach for traffic sign recognition might be, based on the state of the art. The first method consists of extracting HOG descriptors and using an SVM classifier in order to learn the boundary between the different classes of traffic signs, showing better results for larger resolutions. The second one uses a deep convolutional neural network [1] which learns features from the images not much unlike the human brain and eye [2] and the third method uses the more recently developed deep residual network [3], which is currently the best computer vision architecture available. In training the models, the BelgiumTS data set is used, consisting of around 60 classes of traffic signs. We shall observe that newer methods have a much better accuracy, being capable of learning models of great complexity. A comparison between the methods is presented and conclusions are drawn towards the end.

In this paper, we present the benefit of using deep learning time-series analysis techniques in order to reduce computing resource usage, with the final goal of having greener and more sustainable data centers. Modern enterprises and agile ways-of-working have led to a complete revolution of the way that software engineers develop and deploy software, with the proliferation of container-based technology, such as Kubernetes and Docker. Modern systems tend to use up a large amount of resources, even when idle, and intelligent scaling is one of the methods that could be used to prevent waste. We have developed a system for predicting and influencing computer resource usage based on historical data of real production software systems at CERN, allowing us to scale down the number of machines or containers running a certain service during periods that have been identified as idle. The system leverages recurring neural network models in order to accurately predict the future usage of a software system given its past activity. Using the data obtained from conducting several experiments with the forecasted data, we present the potential reductions on the carbon footprint of these computing services, from the perspective of CPU usage. The results show significant improvements to the computing power usage of the service (60% to 80%) as opposed to just keeping machines running or using simple heuristics that do not look too far into the past.

Until recently, CERN had been considered eligible for academic pricing of Microsoft products. Now, along with many other research institutes, CERN has been disqualified from this educational programme and faces a 20 fold increase in license costs. CERN’s current Authentication and Authorization Infrastructure, dating from 2008, comprises multiple Microsoft services from the web Single-Sign-On to the Accounts Database. Replacing these core components is an opportunity to rebuild the CERN infrastructure using the latest technologies and concepts and to respond to evolving requirements of the community. It is also the appropriate moment to consider the alignment of CERN’s and the Worldwide LHC Computing Grid’s approaches to identity management, to create a more consistent environment for operators, developers and users. 2019 saw the launch of an Alpha version of CERN’s next generation Authentication and Authorization Infrastructure, focusing on free and open source products and responding to the limitations experienced by the current system. We describe the new solution and focus on key changes.

Software systems are often maintained by a group of experienced software developers in order to ensure that faults that may bring the system down are less likely. Large turnover in organizations such as CERN makes it important to think of ways of onboarding newcomers on a technical project rapidly. This paper focuses on optimizing the way that people get up-to-speed on the business logic and technologies used on the project by using a knowledge-imbued large language model that is enhanced using domain-specific knowledge from the group or team’s internal documentation. The novelty of this approach is the gathering of all of these different open-source methods for developing a chatbot and using it in an industrial use-case.

Computer security threats have always been a major concern and continue to increase in frequency and complexity. The nature and techniques of the attacks evolve rapidly over time, making their detection more difficult. Therefore the means and tools used to deal with them need to evolve at the same pace if not faster. In this paper the implementation of an Intrusion Detection System (IDS) both at the Network (NIDS) and Host (HIDS) level, used at CERN, is presented. The system is currently processing in real time approximately one TB of data per day, with the final goal of coping with at least 5 TB / day. In order to accomplish this goal at first an infrastructure to collect data from sources such as system logs, web server logs and the NIDS logs has been developed making use of technologies such as Apache Flume and Apache Kafka. Once the data is collected it needs to be processed in search of malicious activity: the data is consumed by Apache Spark jobs which compare in real time this data with known signatures of malicious activities. These are known as Indicators of Compromise (IoC). They are published by many security experts and centralized in a local Malware Information Sharing Platform (MISP) instance. Nonetheless, detecting an intrusion is not enough. There is a need to understand what happened and why. In order to gain knowledge on the context of the detected intrusion the data is also enriched in real time when it is passing through the pipeline. For example, DNS resolution and IP geolocation are applied to it. A system generic enough to process any kind of data in JSON format is enriching the data in order to get additional context of what is happening and finally looking for indicators of compromise to detect possible intrusions, making use of the latest technologies in the Big Data ecosystem.

With the emergence of efficient word embeddings for free text, there has been a bloom of Natural Language Processing (NLP) breakthroughs using deep learning techniques. However, the literature is skewed towards the languages that offer large corpuses, and little research has been done in the direction of Romanian text. In this paper, we propose a Deep Learning (DL)-based system for classifying free sentences in the context of psychological surveys, automatically discovering whether respondees are talking about the expected subject in their answers (thoughts, emotions or behaviors) or not. We test several new architectures of convolutional and recurrent neural networks using pre-trained word embeddings from a very large corpus consisting of Wikipedia and Common Crawl data sets. We also present the benefits of transfer learning applied to NLP, allowing general language models trained on large data sets to be applied to problems that have a small amount of data, as in our case, allowing applications of these techniques in many fields.