Title and abstract of Presentations

Please expand the info on your slot into a record of the form

Time slot:
Speaker:  
Title:  
Keywords:
Short Abstract:

in the following list.

PhD Day Schedule

 October 10

Time slot: 14: 00-14: 30
Speaker:  Federico Tomasi
Title:  Understanding latent variable biological networks through graphical modeling methods
Keywords:  Graphical modeling, network inference, latent variables, biological networks
Short Abstract:  A problem which arises in biological data analysis is that most of the variables which describe one sample may be co-regulated in complicated non-linear interactions. Up to now, very few are the solutions of this problem, since the number of available samples is not enough to capture the interactions among the ever increasing amount of variables. This project focuses on biomedical data structure learning, that is the inference of the possibly complicated interactions between variables at hand. This would be essential in order to understand how the variation of an apparently directly uncorrelated variable affects other variables. Attention towards networks in biology and medicine has been increasing in the last few years but not routinely maintained. My PhD thesis plans to bridge this gap, by establishing and developing a learning method for biomedical data analysis to understand the network structure of data at hand, both static (without temporal information on data) and longitudinal data. Such method would be able to characterise biological variables both with their static (e.g., hubs, subnetworks) and dynamical properties (e.g., evolution in time), by taking into account the complex non-linear interactions that may exist among the variables. Also, I plan to pay special attention to possible non-measured variables (i.e., latent factors), which can be responsible for a large part of measured variables, so being confounding factors.
  

Time slot:  14: 30-15: 00
Speaker:  Luigi Carratino
Title:  Resource Efficient Large Scale Machine Learning
Keywords:  Machine Learning, Kernel Methods, Random Projections
Short Abstract: Kernel methods provide a principled way to perform non-linear, nonparametric learning. They rely on solid functional analytic foundations and enjoy optimal static properties. However, at least in their basic form, they have limited applicability in large scale scenarios due to stringent computational requirements in terms of time and especially memory. In this project, we aim to take a substantial step in scaling up kernel methods to allow them to efficiently process millions of points. The solutions we aim to provide are derived combining several algorithmic principles, namely stochastic subsampling, iterative solvers, preconditioning and asyncronous optimization. The hallmark of the project is the requirement of providing solutions with certifiable and rigorous theoretical guarantees,

Time slot: 15: 00-15: 30
Speaker: Veronica Tozzo
Title:  Graphic modeling of gene regulation development
Keywords: Network analysis, time series, data integration, gene expression, cell differentiation
Short Abstract:At the very heart of life there is the functioning of cells that can be seen as small bricks that form any complex living thing. Each cell contains the same copy of DNA that is able to regulate its functioning in order to form different tissues with different behaviors. Still how from a single embryonic cell an organism is able to differentiate its tissues into a complex body is partially a mystery. Thanks to the now available thousands of measurements from cell, the computational power and the advancement in theoretical modeling we can now extract models of the biological system in analysis from data. In fact the problem is NP-hard and therefore requires specific methods and algorithms to be handled. The most suitable model is the one of a network in which the nodes represent cellular entities and the edges of the interaction between these entities. The way to proceed is reverse engineering of networks from data taking under consideration as noisiness and lack of samples. This thesis proposes to develop computational methods for the inference of biological networks that change over time, caring about the latent processes that we can not measure and the integration of different data sources. The analysis of these networks could give us insights on cell differentiation through gene regulation development in time. This thesis proposes to develop computational methods for the inference of biological networks that change over time, caring about the latent processes that we can not measure and the integration of different data sources. The analysis of these networks could give us insights on cell differentiation through gene regulation development in time. This thesis proposes to develop computational methods for the inference of biological networks that change over time, caring about the latent processes that we can not measure and the integration of different data sources. The analysis of these networks could give us insights on cell differentiation through gene regulation development in time.     

Time slot: 15:30 - 16:00
Speaker:  Gaurvi Goyal
Title:  Action Recognition
Keywords: Action Recognition, Shearlet Transform, Deep Networks, Machine Learning

Short Abstract: Action Modeling and Recognition is a rapidly explored field in the field of human machine interaction. The advances in computational capacities of the system have been a major factor in speeding up this growth. Concentrating on  Action Recognition for Actions of Daily Living (ADL), so far in this PhD, the potentials of spatial temporal primitives developed using the Shearlet Transform have been studied and, at this point, is a promising area of work . Proof of concept exists that the primitives are able to differentiate between actions of considerable difference and also demonstrate a level of view invariance. The next step is to methodically and scientifically explore the potential and extent of the same. On the other hand, Deep Learning Networks have been demonstrated as exceptional tools for the analysis of video data. Therefore, another interesting approach to exploring the applications of Shearlets is the use of Deep Learning to explore the possibility of using the data as being extracted from videos using the Shearlet-based space time primitives as input to a Deep Networks.  

Time slot:  16: 00-16: 30
Speaker:  Angelo Ferrando
Title:  The Book of Trace Expressions
Keywords: Trace Expressions, (Decentralized) Runtime Verification, Multi-Agent Systems, Static Verification
Short Abstract:Runtime Verification (RV) is a software verification technique that complements formal static verification (as model checking) and testing. In RV, dynamic checking of the correct behavior of a system can be performed by a monitor that is generated from a formal specification of the properties to be verified. During these years of PhD, it has been studied the use of the trace expression formalism to achieve the runtime verification of Multi-Agent Systems (and not only). The resulting works are mainly focused on innovative aspects such as the extension of formalism to represent more complex properties, the decentralization of the runtime verification process and the hybrid combinations of static and dynamic verification (taking the best of them). In the presentation will be summarised the major results obtained this year and will be presented the next steps scheduled as objectives for the last year of the PhD.

Time slot:  16: 30-17: 00 
Speaker:  Vanessa D'Amario 
Title:  Understanding neural networks through decoding for pattern identification
Keywords: Machine learning, signal processing, neural networks, deep learning, information theory, neuroscience
Short Abstract: Information theory and machine learning are different and complementary fields useful in the analysis of temporal data. The application of the former to the signal analysis, through binary coding and decoding techniques is used mainly in the telecommunication field for transmission. It could provide an interesting starting point for a better comprehension of data coming from noisy sources. Its peculiar characteristics directly lead  to profound questions about possible analogies with very popular machine learning methods, such as Deep Learning. I will present open questions about the plausible links between the two fields in the context of time series analysis, with focus on neural activity patterns.

Time slot:  17: 00-17: 30 
Speaker:  Francesco De Fino 
Title:  Exploiting Recurring Retrieval Needs in Querying Heterogeneous and Dynamic Graph Dataspaces
Keywords: Data management, Query processing, Graph query, Recurring retrieval need
Short Abstract:  Querying data sources containing heterogeneous and dynamic information is a complex task: high quality answers need to be produced fast to cope with the dynamicity of the environment. A common solution to reduce query processingtime is to rely on approximate processing approaches, which provide a faster answer at the cost of a lower accuracy. On the other hand, user involvement in the interpretation of the request, is not adequate in dynamic contexts, where urgent requests hamper user interaction. In order to cope with the difficulties raised by the heterogeneity and dynamic nature of the considered environments, it has been claimed that new solutions should be devised, relying either on additional knowledge about the current execution (e.g., query context or user profile) or previous executions of similar requests. The goal of the thesis is to exploit information related to requests recurring over time for efficiently and effectively process, in an approximate way, complex requests on heterogeneous and dynamic data spaces, possibly available on a highly distributed environment, while guaranteeing user satisfaction and limiting as much as possible user interactions. The idea is to take advantage of prior processing in order to obtain shortcuts to different points in the query processing stack. The approach we envision generalizes smart caching methods allowing various kinds of information to be associated with cached queries in order to improve query processing even further. More precisely, the thesis will first propose a general framework for recurring queries representation and management, as well as two specific instantiations, related to two particular application contexts, namely source selection for linked data and mapping selection for pay-as-you-go data integration. The role of request contexts and data quality in developing such solutions will also be investigated.

   
October 12

Time slot: 14:00 - 14:30
Speaker:  Diego Clerissi
Title:  An overview of web applications and IoT systems acceptance testing
Keywords: Acceptance Testing, Requirements Engineering, Web Applications, IoT systems
Short Abstract:
Web applications pervade our life and are crucial for a multitude of activities, thus their quality has become a top-priority problem. End-to-End testing exercises a web application as a whole, by adopting its requirements specification as a reference for the expected behaviour. A good solution in testing a web application is relying on test automation practices, where the manual effort is reduced as much as possible. Another challenge to face nowadays is testing Internet of Things (IoT) systems, where interconnected physical objects and devices use the Internet to communicate in a safety-critical condition.
In the last few years, several approaches and techniques have been proposed for automated testing of web applications and IoT systems, but still few works have been conducted towards acceptance testing of IoT systems and in actual studies of requirements and test artefacts relationships. In this talk, an overview of some of the proposed approaches is given.

Time slot: 14:30-15:00
Speaker:  Alberto Cabri
Title:  Computational Intelligence Methods for Massive Data Learning
Keywords: machine learning, sequential classification, deep learning, big data, early decision taking
Short Abstract: The ever growing amount of data that becomes available thanks to the diffusion of connected devices and applications inevitably affects our ability to analyze them to infer their semantics, always requiring new approaches, capable of taking into account the three main V’s of Big Data: Volume, Velocity and Variety. A key point in data analysis is definitely the speed of data creation, that may represent a competitive advantage in business as it becomes even more important to be able to process both structured and unstructured data as a stream and take actions according to data dynamics, highlighting the need for reliable early decision making algorithms. The project is focused on studying and testing the effectiveness of deep learning methods in new research areas that imply both early decision making and real time data transformation and analysis, with high potential impact on health and business applications.

Time slot: 15:00-15:30
Speaker: Tommaso Petrucciani
Title: Set-theoretic types for functional and object-oriented languages
Keywords: type systems, union and intersection types, semantic subtyping, gradual typing
Short Abstract: Static type systems are valuable tools to check programs for errors, but they can also be restrictive and cumbersome to use. Much research therefore strives to define more expressive and less intrusive systems. The theory of set-theoretic types (union, intersection, and difference types) and semantic subtyping is one way to type programs expressively while maintaining the system fairly simple to use. During my years of PhD, I have worked on trying to expand the applicability of set-theoretic type systems to language features including type inference, polymorphic record types, lazy evaluation, and gradual typing (an approach to combine static and dynamic typing in the same program).

Time slot: 15:30 - 16:00
Speaker: Luca Franceschini
Title: Trace Expressions for Runtime Verification of Node.js Systems
Keywords: Runtime Verification, Trace Expressions, Node.js, JavaScript, Internet of Things
Short Abstract: Node.js is an environment that allows JavaScript code to be executed outside a browser. It gained popularity as it allows developers to employ JavaScript for server-side programming, and more recently, for the Internet of Things. However, the dynamic nature of the language, together with the non-determinism of Node.js applications, makes it extremely challenging to ensure correct behavior with traditional approaches like static/formal methods and testing.
Runtime verification is an alternative software analysis approach in which a running system is observed by a monitor that perceive relevant events and verifies their correctness against a given (formal) specification of the expected behavior. We propose the use of a formalism called trace expressions to specify the expected properties, and based on this we implemented a prototype of a monitoring system for Node.js applications.
The aim of this work is to study the use of trace expressions in the context of runtime verification and to build a complete framework for using them to verify Node.js code in real-time, automatically generating the required components from the formal specification.

Time slot: 16:00-16:30
Speaker:  Simone Aonzo
Title:  Improving the reliability of Android for malware detection
Keywords: android, malware, mobile security
Short Abstract: Several static and dynamic analysis techniques were developed in recent years to detect and analyze Android malware. However, all of them are weak against specific malware obfuscation mechanisms. For instance, static analysis techniques are useless when malware exploit Java reflection and JNI to hide their behavior. Furthermore, dynamic analysis techniques reveal to be rather weak against malware that implement emulator detection and integrity self-checking as such mechanisms allow the malware to detect and bypass all recent Android sandboxes. In this scenario, the aim of this thesis is to put forward a modification of the Linux Kernel at the basis of the Android OS, that would allow to expose a set of APIs to trusted Android applications. Such APIs would allow trusted applications to monitor the behavior of other applications at runtime, in order to detect when malware obfuscation mechanisms are used.

Time Slot: 16.30 – 17.00
Speaker: Bassano Chiara
Title: Natural interaction in an immersive virtual environment
Key words: Virtual Reality, Sense of Presence, Natural Interaction
Short abstract: Virtual Reality (VR) is having a widespread success, thanks to the recent release of low cost technologies. These devices are able to provide immersive experiences and intuitive and effective interfaces. Moreover, the headsets are sold in combination with controllers and simple tracking systems. This way, the user can walk in the virtual world and interact with it without the need of sophisticate and expensive motion capture systems. However, there are some well-known issues related to VR: the problem of depth underestimation; the absence of a proper representation of user's body; the use of tools to interact with VR, which are stable and effective but could break the illusion of being in VR.

The aim of my work is investigating all these aspects in order to obtain a more realistic experience of VR, based on the naturalness and intuitiveness of the interaction and on the multimodality, which is the involvement of different sensorial channels (visual, proprioceptive, vestibular and even haptic).