Mauriana Pesaresi

Modern applications are developed into sets of hundreds of interacting microservices that implement one business logic. Such a result is achieved through a collaborative, incremental and iterative development process. Furthermore, the Cloud paradigm has shown some drawbacks in supporting new-generation microservice-based applications. To tame such limitations, Cloud-Edge computing paradigms emerged to bring computation closer to end-users, by exploiting computing, storage, and networking resources along a multi-layered continuum from the Cloud, through Edge, to IoT nodes. In this context, microservices are to be orchestrated over large-scale, pervasive, heterogeneous Cloud-Edge resources to meet application requirements continuously. The constant infrastructural changes and the iterative development process require scalable and continuous orchestration capabilities to manage application deployments and guarantee all requirements by promptly reacting to infrastructure or application changes at runtime. Thus in this talk, we analyse and discuss the key factors and problems arising in the orchestration of multiservice applications in Cloud-Edge infrastructures.

Quantum computing is a promising emerging technology, that exploits non-classical phenomena described by quantum mechanics, such as entanglement and superposition. Quantum algorithms are predicted to have a exponential speedup with respect to their classical counterparts. Due to technological limitations, however, we still have not reached “Quantum Supremacy”, the point where a quantum computer is undeniably faster than classical ones. Larger and more powerful quantum computer are being built each year, but at the same time faster and smarter algorithm to simulate them are being developed. It is still an open question when (or even if) quantum computers will provide a computational speedup, and which are the real world problems they will benefit the most.

The Internet of Things (IoT) has been revolutionizing the way we live and act in recent years. The way these devices can act in different environments, gathering data, providing services, allowing interactions, has brought several possibilities in the field of research. Not so long ago, it was unimaginable how popular these devices would become, in the most varied forms, from smart bands and smartwatches that monitor sleep or our daily physical activities to controllers for smart homes or industry 4.0. The extensive use of IoT demands light and scalable solutions, with guarantees of security and privacy. The concept of blockchain, has great potential for integration due to its distributed, secure, and private nature. We can say that blockchain has opened the door to the development of an open, scalable digital economy and without the need for third-party supervision to make transactions. Other cryptocurrencies with similar technology have emerged since the first system using blockchain was released, with the Bitcoin cryptocurrency. Within the context presented so far, the Economy of Things (EoT) paradigm emerged. We can define EoT as the monetization of things, or the possibility to exchange digital assets. The integration between blockchains and IoT can bring many advantages to this scenario.

The Internet of Vehicles (IoV) is an emerging concept that refers to the integration of vehicles with the internet, road units, and other vehicles. The IoV promises to revolutionize the automotive industry by enhancing the efficiency and safety of vehicles. However, with the increasing connectivity of vehicles, the threat of cyber-attacks on the IoV has become a significant concern. Cybersecurity is a critical issue in the IoV as any potential breach can have severe consequences for the safety of drivers and passengers. Cyber threats to the IoV can include hacking, data theft, and denial of service attacks. In this talk, we describe the IoV and its main features with a focus on in-vehicle networks. Then, we discuss vehicle security with examples of some attacks that a car can suffer. In particular, we would like to find some possible ideas and solutions to increase the security and privacy of IoV.

Protein-protein interactions (PPIs) describe the physical contacts between two or more proteins and are crucial for a wide range of biological processes, including cell signaling and metabolic and developmental control. The study of PPIs is of pivotal importance for understanding molecular mechanisms, drug design, and the study of mutations. In recent years, many computational PPI prediction techniques have been developed to compensate for the excessive costs and technical limitations of wet-lab experiments (e.g., Deep Learning (DL)). Using DL for PPIs can help design new molecules (e.g., antibodies) with specific structural and functional requirements. This seminar will highlight the main issues in antibody-antigen interaction prediction and de-novo antibody design.

Dealing with normalization requirements mostly comes with untractable hurdles to solve. We will briefly recap recent intuitions behind score-matching methods, which allowed to overcome these issues computationally and in terms of results.

In recent years, Reinforcement Learning has reached astonishing results and advancements in various research fields and applications such as games, robotics and chemistry. However, this has come at the cost of unsustainable computational efforts given the huge and complex deep architectures. A common characteristic of all state of art approaches is that the agent learns to solve a very specific task without reusing previously learned skills or knowledge. Continual Learning, on the other hand, seeks to create models that are able to incrementally learn from a sequence of tasks, while continually acquiring, fine-tuning, and transferring knowledge and skills throughout different tasks. As each task is temporally limited and discarded, the ability to forget and maintain knowledge is a critical and fundamental aspect. Continual Reinforcement Learning aims at creating agents that learn multiple tasks in a sequential fashion, trying to extend their abilities without loss of performances in previously learnt scenarios. But at the same time, agents are able to re-use or transfer previous knowledge to ease the process of learning a new task. In this presentation, I will cover the current state-of-the-art and analyze possible ideas that may help to improve the agent's performances and capabilities.

Countries around the globe are working towards reducing their carbon emissions and moving towards clean energy. This is particularly visible in the surge of e-mobility over the past few years. A driver's decision to purchase an electric vehicle is influenced by factors such as the desire to phase out vehicles that make use of fossil fuels due to climate change coupled with the increasing fuel and gas prices, the availability of sufficient charging infrastructure, and the charging speed. The insufficient availability of charging infrastructure and long charging times may lead to range anxiety and discourage drivers from switching to electric vehicles.

With the increased use of mobile devices, which capture information from users' locations using GPS and /3G/4G hotspots/cells, the trajectory data created from their spatio-temporal evolution has grown considerably. Furthermore, with the advent of the Internet of Things (IoT), popularization of social networks and mobile devices, a massive amount of data has been produced every single second. This data can be generated by smart vehicles, houses, sensors, watches, appliances, as well as social networks (e.g. Facebook, Instagram, Twitter, Linkedin) and location-based services. We can use this data from these sources to enrich the trajectory components. They can be related to the object (e.g. heartbeat, mood, blood pressure, sleep, age, occupation, gender) or the location (e.g. description, temperature, air and noise pollution, category, assessments). This diversified type of data has its characteristics and, when combined to trajectories, creates a new intricate trajectory model. This complex and heterogeneous trajectory that contains numerous dimensions is known as multiple aspects trajectory. This data can be useful for several areas of knowledge, including security, urban planning, public transport management, and epidemic prevention. However, the publication of such data may put the privacy of users at risk. Stalking can be facilitated if malicious persons have access to this data, as well as operational support for committing crimes, leading to a threat to the people's safety. In this seminar, we will talk about the current challenges related to privacy and multiple aspects trajectories.

Continual self-supervised learning for video action recognition is an open problem that has garnered significant attention in the field of computer vision. The ability to learn from unlabeled data is of particular interest, as it has the potential to reduce the cost and effort associated with labeled data collection. However, current self-supervised learning methods for video action recognition have limitations when it comes to continual learning. Specifically, these methods are typically designed to learn from a fixed set of data, rather than adapting to new and evolving scenarios over time. To address this challenge, researchers have begun exploring novel approaches to continual self-supervised learning for video action recognition. These approaches include using meta-learning techniques to learn how to learn from new data, building models that can adapt to new and changing data distributions, and developing methods that can incorporate human feedback to improve learning over time. While these approaches show promise, there are still significant challenges to overcome in the field of continual self-supervised learning for video action recognition. One such challenge is the lack of benchmark datasets that are specifically designed for continual learning. Another challenge is the need to develop effective methods for balancing the trade-off between exploiting existing knowledge and exploring new data. Despite these challenges, continual self-supervised learning for video action recognition remains an important area of research with the potential to revolutionize the field of computer vision. By developing novel methods and approaches to this problem, researchers can unlock the power of self-supervised learning to create more efficient and effective models for video action recognition.

It is evident that digitalization and automation, bought a lot of business benefits but modern industrial systems require more efficient, trusted, and secured ways of communication. The trend of developing innovative and smart architectures notably in the fields of manufacturing, logistics, healthcare, and agriculture. To build these smart systems, the utilization of more IoT devices is increasing day by day. Industry leaders want to define clear business models and processes, which are reusable reliable reproducible and can be monitored adequately. Industry 4.0 provides connectivity and interconnectivity among industrial systems which has increased the number of IoT devices with broader challenges. It is paramount to mention that IoT ecosystem is supported by cloud solutions and data analytics to extract appropriate information for making our decisions. In this type of systems cyber-security becomes vital to have secure communication. In recent time Google cloud Microsoft Azure and Amazon are the most established cloud platforms because of their easy integration with IoT. Despite being world best robust, trusted, reliable commercial solutions the number of IoT devices are increasing in their system and indirectly it's expanding the surface of cyber-security threats and attacks. However, without secure communication It becomes quite difficult to use IoT systems.

Mechanical metamaterials are a category of human-made objects engineered to exhibit specific responses when subjected to mechanical forces. Their behavior is determined by the arrangement of internal components, primarily based on their geometry. These materials serve as fundamental units that can be combined to construct larger mechanical structures, resulting in entities that possess both macroscopic (overall shape) and mesoscopic (internal structure) geometries. By leveraging mechanical metamaterials, designers can explore novel design possibilities beyond what traditional techniques offer. However, this shift also necessitates the development of new computational tools tailored to address the design challenges at the mesoscale.

It is an obvious fact that every evolution in HPC systems, will bring new complexities in applications workloads, due to increment in number of elements that compose the workflows, and also the type of computations they perform. While simulations and modelling of physical phenomena are targeted in traditional HPC workflows, there are some problems that require extra task computation in artificial intelligence (AI) and data analytics (DA). However, development of these workflows is prevented, mainly because of absence of appropriate programming model and environment that assist integration of HPC, AI, and DA, plus lack of tools to smoothly deploy and execute workloads in HPC platforms. To proceed in this way, use cases are presented where complex workloads are required and main issues that need to be addressed in HPC/DA/IA convergence is investigated. For doing so, first a software stack that provides the functionalities to manage these complex workflows is defined, and second, HPC workflows as a Service (HPCWaas) paradigm is proposed, which leverages the software stack to facilitate the reusability of complex workflows in federated HPC infrastructures

Since John von Neumann and Oskar Morgenstern published the book Theory of Games and Economic Behavior in 1944, modern game theory has been developed and applied in many fields such as economics, social science and also computer science. In which, coalition game is a family of games that draws a lot of attention. The theory of coalition game provides a high-level approach to model many problems. In this seminar, we will briefly review coalition game and the solution proposed by Shapley for this game together with its application in several current issues such as Machine Learning, Computational Biology.