Mauriana Pesaresi Seminars 2018

Mauriana Pesaresi Seminars 2018 Speaking Calendar

27 February 2018
Speaker: Oleksii Osliak
Title: Towards to privacy-preserving data analysis of structured CTI
Abstract: Cyber Threat Intelligence (CTI) plays one of the most significant roles in cybersecurity. By sharing this information, organizations can improve their cyber threat situational awareness. As a result, shared CTI helps in developing a better course of necessary actions in order to react faster and more effective against new malicious activities. However, it has many challenges and limitations. This work aims to address one of them by proposing a possible solution.
06 March 2018
Speaker: Lorenzo De Mattei
Title: Multi-modal and Multi-task learning for NLP
Abstract: In Machine Learning, we typically care about optimizing for a particular metric, whether this is a score on a certain benchmark or a business KPI. In order to do this, we generally train a single model or an ensemble of models to perform our desired task. We then fine-tune and tweak these models until their performance no longer increases. While we can generally achieve acceptable performance this way, by being laser-focused on our single task, we ignore information that might help us do even better on the metric we care about. Specifically, this information comes from the training signals of related tasks. By sharing representations between related tasks, we can enable our model to generalize better on our original task. This approach is called Multi-Task Learning (MTL). Moreover the information in real world usually comes as different modalities. For example, images are usually associated with tags and text explanations; texts contain images to more clearly express ideas, videos usually are associated with audio ecc. Due to the distinct statistical properties of different information resources, it is crucial to let machine learning system to take advantage from all the different source of information. In this talk I will introduce my PhD research project that aims to (i) discovering the impact of Multi-task and Multi-modal learning in different NLP tasks and, from a more theoretical point of view, (ii) understanding the impact of multi-modality and multi-tasking from a cognitive point of view.
20 March 2018
Speaker: Fabrizio Luccio
Title: Chip functioning and manufacturing
Abstract: Half a century ago Gordon E. Moore, in a four page article published in a trade magazine, predicted that all the circuits of computers, mobile phones, and other control systems would be regulated by a yearly doubling in the number of components that can be economically packed in an integrated circuit. The growing rate of this prediction has been modified over the years, but an exponential growth has been essentially observed up to today. We discuss some facts that have made, and will still make this possible. After an overview of MOS technology that has dominated digital systems up to now, we present the late innovations that are transforming the field, and some of the future directions that chip fabrication may take. This is not a research talk, being motivated by my impression that these topics are not treated in depth in the informatics curricula. If you already know the field, please use your time more fruitfully.
27 March 2018
Speaker: Luca Rinaldi
Title: Orchestrate incomplete application with TOSCA and Docker
Abstract: How to package deploy and manage complex applications across heterogeneous cloud platforms is one of the main concerns in today's enterprise IT. Cloud applications typically integrate multiple components, each needing a virtualised runtime environment that provides the required software support (e.g., operating system, libraries). This presentation shows a close look at TosKeriser and TosKer, which synergically combine TOSCA and Docker to effectively support the orchestration of multicomponent applications, even when their runtime specification is incomplete.
22 May 2018
Speaker: Antonio Carta
Title: Linear Memory Networks
Abstract: Recurrent neural networks can learn complex transduction problems that require maintaining and actively exploiting a memory of their inputs. Such models traditionally consider memory and input-output functionalities indissolubly entangled. We introduce a novel recurrent architecture based on the conceptual separation between the functional input-output transformation and the memory mechanism, showing how they can be implemented through different neural components. By building on such conceptualization, we introduce the Linear Memory Network, a recurrent model comprising a feedforward neural network, realizing the non-linear functional transformation, and a linear autoencoder for sequences[1], implementing the memory component. The resulting architecture can be efficiently trained by building on closed-form solutions to linear optimization problems. Further, by exploiting equivalence results between feedforward and recurrent neural networks[2] we devise a pretraining schema for the proposed architecture. Experiments on polyphonic music datasets[3] show competitive results against gated recurrent networks and other state of the art models.
29 May 2018
Speaker: Francesco Crecchi
Title: On the state of Adversarial Machine Learning
Abstract: Since 2013, with the rise of deep learning, it is unusual to think that “AI” based programs may fail. Indeed, researchers have found that this is the case with Adversarial Examples: carefully crafted examples to cause a misclassification. This field, so-called Adversarial Machine Learning, is facing an arms race between attackers and defenders of machine learning based models, leading to an exponential growth in scientific productions during the last four years. This presentation aims to give you a general introduction to the problem of Adversarial Examples and on the most powerful attacks/countermeasures that have been deployed so far.
5 June 2018
Speaker: Matteo Busi
Title: An Algorithmic Schema for Incremental Type Checking
Abstract: Type checking is a widespread technique to avoid runtime errors in programs. However, the ever-growing size of programs and the adoption of new development models that prescribe a continuous evolution of the code require building fast and efficient type checkers. A promising solution to this problem is to make type checkers incremental so that they do not need to consider the entire codebase but just those parts that changed and what depends on them. During the talk, I will show an algorithmic schema that drives the definition of an incremental type checker and I will illustrate how to instantiate it to define an incremental type-checker for an imperative language.
12 June 2018
Speaker: Andrea Michienzi
Title: Exploiting dynamic network properties to improve complex systems
Abstract: By now complex network analysis is a widely used tool to model and study large scale phenomena in a number of fields of research: from biology to economics, from social sciences to computer science and many more. The overwhelming majority of tools, and therefore of studies, focus networks from a purely static (non-changing) point of view. While in some areas this simplification is reasonable, in many other is not. Consider, for instance, the social relations among people: they change over time and are influenced by a number of factors, some of which are unpredictable. Thus, it is necessary to study such problems in a dynamic fashion, observing how these networks change over time, to be able to improve the systems behind. As a case study, we will see the hot topic of Online Social Network (OSN). OSNs are platforms that let users create a personal profile, and produce, consume and share contents on the platform. The service is generally speaking free, but the providers require users to accept special terms of service that allow user's data to be analysed and, eventually, sold for marketing purposes. This has arose a lot of concern, especially from common users. In order to give users an higher sense of privacy, scientists have developed Decentralized Online Social Networks (DOSN) which are OSNs implemented on decentralized architectures. The decentralization offers some advantages, besides more privacy to users, but also introduces a number of new problems to be addressed. In this presentation we will see that dynamic network analysis tools can be used to define new flexible and powerful strategies to tackle these problems.
19 June 2018
Speaker: Mohsin Ur Rahman
Title: Lightweight detection of malicious nodes in mobile ad hoc networks
Abstract: Mobile Ad Hoc Networks (MANETs) are self-arranging, dynamic, infrastructure-less and multi-hop networks formed by a group(s) of mobile nodes (MNs), which transmit information via wireless channels. The broadcast nature of the wireless medium makes MANETs vulnerable to various types of attacks (e.g., masquerading attack, node replication attack, Sybil attack and so on). A critical challenge that allows replicas to defeat the well-known detection algorithms is known as abnormal or selected silence. This problem arises when replicas intentionally become silent at specific time periods or stop broadcasting detection packets. As a result, the attackers become invisible and successfully evade the detection process. Most of the existing detection techniques are vulnerable to these malicious activities. In this paper, we propose a new lightweight technique for mitigating the above issue in MANETs. Our solution is lightweight and does not need the use of extra communication or geographical positioning system (GPS). Simulation results show excellent detection accuracy of the proposed scheme.
3 July 2018
Speaker: Luca Versari
Title: Algorithms for Analyzing Massive Networks
Abstract: Network analytics provides algorithms for identifying trends and patterns in networked data and social networks, with massive data sets produced at an ever increasing pace by industry, science, and people. This talk describes new powerful ideas to design algorithms for gaining deep insights in many critical tasks for business and scientific research, such as community detection. These algorithms scale well with massive real-world graphs, which exhibit common properties (e.g., sparse, clustered), and are able to process datasets which are orders of magnitude larger than previous approaches.
12 July 2018
Speaker: Adrian Spătaru.
Title: Exposing HPC-aware Cloud Services.
Abstract: Cloud Service Providers offer Virtual Machines that are able to run services, which mostly comprise of web applications. More recently, they started to offer support for GPUs and FPGAs, but is the consumer's job to choose the software stack and parameters to execute an application making use of the accelerated compute units. This talk will present the CloudLightning UI and Service Description Language, through which a cloud consumer can use services like RayTracing (rendering), Oil and Gas Simulation, and Genomics Alignment in the form of Docker containers leveraging HPC infrastructure, or simple VMs.