Link to doodle for managing the schedule
Link to doodle for managing the pizza orders
Mauriana Pesaresi Seminars 2019 Speaking Calendar
This is the page for the Mauriana Pesaresi Seminar of 2019. Here there will be infos regarding the spekers and their topics.
- 13 March 2019
Speaker: Stefano Chessa
Title: Scientific papers and their performances
Abstract: The talk shares considerations and observations concerning the quality and performance of research, accrued by the speaker during his activities as researcher and in support to the activities of evaluation of the quality of research in his own department.
The talk introduces the main performance indicators and it discusses the reasons why some papers achieve performances higher than other papers, by presenting case studies taken from the speaker's own experience. The last part of the talk discusses some bad practices and explains why they are not a good idea.
The presentation is geared primarily to doctoral students, but it is, of course, open to everybody.
26 March 2019
Speaker: Irene Sucameli
Title: Describing verbs using visual vectors
Abstract: In the research field of NL Processing and Understanding, perhaps one of the most challenging and interesting task is describing human knowledge and understanding. With this research we will try to answer to the following question: is it possible to define the semantic similarity between verbs using images? Our approach uses visual distributional semantic models applied to verb analysis. Moreover, our research lays the groundwork for the development of future a multimodal DSM.
2 April 2019
Speaker: Alessio Miaschi
Title: Natural Language Processing in the educational environment
Abstract: The educational environment is changing on a drastic speed,
from traditional classroom teaching ecology from the adaptive
individual/collaborative learning. In this scenario, the interest in
applying Natural Language Processing to education has rapidly increased.
The talk discusses the main challenges in applying NLP to education,
focusing principally on two research projects: tracking the evolution of
student's writing skills and processing text from the web to personalize
16 April 2019
Speaker: Lorenzo Ceragioli
Title: High Level Management of Firewall Configurations
Abstract: Firewalls are one of the standard mechanisms for protecting computer networks, the correctness of their configuration is of critical interest.
Unfortunately, managing firewall configurations is notoriously an hard and error prone task, mainly because firewall languages are rather complex, low level and don't have a clear semantics.
We present a formally-ground language-based approach for generating, analysing, porting and refactoring configurations at high level; it is based on a transcompilation pipeline between firewall languages, and implemented in a tool called FireWall Synthesizer (FWS).
We also present a function-based redefinition of this pipeline, and some ongoing works devoted to fully support tag systems, and allow high level management of networks of firewalls.
30 April 2019
Speaker: Riccardo Guidotti
Title: Designing Clustering Algorithms For Personal Data Analytics
Abstract: Mining a large number of datasets recording human activities
for making sense of individual data is the key enabler of a new wave of
personalized knowledge-based services. We focus on the problems of (i)
clustering individual transactional data and (ii) clustering stop points
for a large mass of users. We present TOSCA and TX-Means, two
parameter-free clustering algorithms able to efficiently partitioning
transactional data and mobility data, respectively, in a completely
automatic way. A deep experimentation on both real and synthetic
datasets shows the practical effectiveness of the proposed algorithms
and proves that they outperform existing methods in terms of quality and
efficiency. We also present applications of them in real systems.
7 May 2019
Speaker: Mattia Setzu
Title: Tackling Explainability in Machine Learning
Abstract: Machine learning is part and parcel of our society just as any other prominent computational field, and we expect it to grow more pervasive over time. Its impact on society is undeniably multifaceted and raises, among others, great concerns about its fairness and transparency. Among others, Explainability addresses these issues by providing an interpretable layer on top of trained opaque machine learning models. In this talk we will introduce the aforementioned research field and ETHICA, a "local-first" model agnostic explanation algorithm. We will highlight its strength and weaknesses and present a plethora of empirical analysis on its performance. In conclusion, we will present some current research directions involving Latent Spaces and Counterfactual generation.
14 May 2019
Speaker: Giorgio Vinciguerra
Title: Superseding Traditional Indexes with Multicriteria Data Structures
Abstract: The ever-growing amount of information coming from the Web, social networks and IoT severely impairs the management of data. Much research has been devoted to dealing with this issue, however, we still miss proper algorithmic solutions that work under computational constraints that vary across users, devices and time.
We therefore propose the concept of Multicriteria Data Structures, which add to the classic requirement of being efficient, the novel feature of dynamically adapting themselves to the constraints imposed by the application of use.
We show the potentiality of this concept by focusing on the paradigmatic dictionary problem. After reviewing some old and new solutions to this problem, we present the first multicriteria data structure, the PGM-index, that solves it with better (both asymptotically and experimentally) time and space than classic indexes for hierarchical memories, such as B-trees, and modern learned indexes. Finally, we show an optimisation algorithm that finds the best design setting of a PGM-index according to the input constraints (either in time or in space).
7 June 2019
Speaker: Benedikt Bienhüls
Title: Service Planning and Disruption Management in Public Transport Systems
Abstract: Disruption management can be at least as challenging as the problem faced during the planning phase of a public transport system (that is, if a disruption occurs at the beginning of the service),which consist of multiple NP-hard subproblems. Solution approaches in the literature can be separated into sequential, semi-integrated and fully-integrated ones and range from heuristics to exact methods. We start with an overview of the steps involved in public transport planning and disruption management. Afterwards we explain some mathematical concepts used to tackle the problems involved. Finally, we look at methods that have been used in the literature and present different ways of approaching the problem.
11 June 2019
Speaker: Giuseppe Attardi
Title: The Tsunami of Deep Learning on NLP: the sequel
Abstract: Five years ago I gave an invited speech with this title at the Italian conference on Artificial Intelligence.
I presented new techniques that were drastically changing the approaches to NLP.
Recently, new advances in the field are making those approaches to be considered “old NLP”.
It all started with the introduction of a mechanism of attention for sequence to sequence models used in neural machine translation. This was so effective to lead to suggest that “All you need is attention”, in the paper that introduced a general purpose language model, which accumulates knowledge of languages from billions of sentences.
A single such model can be fine tuned to perform a variety of tasks, improving the SotA in all of them.
Building such models requires significant computing resources, but they are so good that OpenAI restrained from distributing its model based on ethical considerations. The largest models require larger GPUs than one can normally afford, also just for tuning.
The new tsunami is therefore both exciting and dangerous.
Finally I will suggest four scientific challenges for current AI research.
25 June 2019
Speaker: Fabrizio Luccio
Title: Simple mathematical proofs that you may not have seen in school
Abstract: The history of Mathematics is fill of intriguing results that are not always taught in school, some of which are amazing for their same nature and for relying on important proof methods. Among countless cases I will stay with seven. Pythagoras theorem and rigid trasformations. Irrational numbers and reductio ad absurdum. Fibonacci numbers and power series. Euler's theorem on polyhedra and an invariant-preserving algorithm. A paradox from the school of Galileo and integral calculus. The ellipse, simmetry and a metric argument. Amazing sequences and potential functions. This talk is for recent members of the department, as it is an improved remake of one given several years ago in this same series.
2 July 2019
Speaker: Vahid Zolfaghari
Title: Security analysis of Internet of Things considering the Social relationships
Abstract: Traditionally, network traffic monitoring tools have focused merely on network-oriented metrics such as volume of data exchanged or top host talkers. Recent cybersecurity attacks instead demonstrated that social relationships have a great impact on network threats. These attacks exploit social relationships such as a shared disk between friends or people belonging to the same working group. To contrast cybersecurity attacks of this kind, novel analysis techniques need to be developed, which do not focus exclusively on packet-level analysis, but correlate traffic patterns with the properties of the nodes generating them (e.g., the same traffic pattern might be legitimate or not, depending on whether the communicating endpoints belong to the same user, to members of the same social community, or to complete strangers).