Link to doodle for managing the schedule
Mauriana Pesaresi Seminars 2020 Speaking Calendar
This is the page for the Mauriana Pesaresi Seminar of 2020. Here there will be infos regarding the spekers and their topics.
- 31 January 2020 [11.00 - 13.00]
Speaker: Beppe Attardi and Elisabetta Savigni Ullmann
Title: Lost in translation, human and non-human interpreters
Abstract: The two speakers will comptete on the topic whether or not human interpretation and translation should/could be replaced by automatic ones.
Professor Attardi will tell us recent AI results which have made possible, in the field of automatic interpretation and translation,
achievements unthinkable only a few years ago. Doctor Ullmann will tell us, through numerous anecdotes from her long career, aspects of interpretation and translation
that are very difficult or impossible to automate.
- 10 February 2020 [13:00 - 14:00]
Speaker: Gabriele Lagani and Giuseppe Amato
Title: Hebbian Learning Algorithm for Training Convolutional Neural Networks
Abstract: Learning algorithms based on Gradient Descent and error backpropagation are very popular for training Artificial Neural Networks (ANNs).
Despite the great success, these approaches are considered to be implausible from a biological point of view.
In this presentation, we are going to explore biologically plausible alternatives for neural network training based on the Hebbian principle:
"Neurons that fire together wire together".
In particular, we are going to see how one such alternative, based on a form of interaction among neurons known as Winner-Takes-All (WTA) competition,
can be used for training deep Convolutional Neural Networks (CNNs) on computer vision tasks.
- 17 February 2020 [13:00 - 14:00]
Speaker: Salvatore Citraro and Giulio Rossetti
Title: Mixing pattern quantification in node-attributed networks
Abstract: Unveil the homophilic/heterophilic behaviors that characterize the wiring patterns of complex networks is an important task in Social Network Analysis.
Recent works underlined that a global measure necessarily provides partial picture of the reality.
Moving from such literature, we are working on a novel node-centric measure, namely Conformity, designed to overcome such limitation.
Conformity is designed to be path-aware, allowing for a more detailed evaluation of the impact
that nodes at different degrees of separations have on the homophilic embeddedness of a target.
We will present preliminary experimental analysis and a possible statistically significant framework of comparison between Conformity
and another state-of-art measure, namely Peel's assortativity, in the absence of a ground truth that allows a direct comparison of the two formulas.
- 24 February 2020 [13:00 - 14:00]
Speaker: Daniele Di Sarli
Title: Exploiting Randomness in Neural Networks
Abstract: Nowadays, Machine Learning is often associated with long and costly training times.
Neural network models from many recent works have billions of parameters that have to be fine-tuned by learning algorithms on huge amounts of data.
First, in this presentation we are going to introduce the field of Reservoir Computing, an alternative approach to neural network training
which takes inspiration from dynamical system theory and biological neural networks.
Then, we are going to present the innovative application of Reservoir Computing techniques to the field of Natural Language Processing
with the aim of drastically reducing training times.
- 2 March 2020 [14:00 - 15:00]
Speaker: Andrea Lisi and Paolo Mori
Title: Distributed Ledger Technology: An introduction to Interoperability
Abstract: Thanks to Distributed Ledger Technology (DLT) is possible to build decentralized state machines which agree on the evolution of a set of data.
For example, the Bitcoin protocol exploits a particular DLT, the Blockchain, to maintain a decentralized payment system.
However, protocols based on DLT tend to be self-contained systems with no access to external information without any third party system.
This presentation provides an overview on the most famous DLT protocols, Bitcoin and Ethereum,
and on the current state of the art concerning DLT interoperability models and use cases.
From 16 March Pesaresi's seminars will continue on Microsoft Teams. The code for joining the team is djikfzt
- 16 March 2020 [14:00 - 15:00]
Speaker: Giacomo Iadarola
Title: Designing Robust Software Analysis and Artificial Intelligence Approaches For Cybersecurity
Abstract: Smartphones, laptops and smart devices (the Internet-of-Things) brought all kind of technologies into our homes and makes them accessible to everyone. Being a friendly piece of technology means collecting a huge amount of sensitive and personal information, besides becoming an attractive treasure for attackers that want to profit from users data.
Thousands of applications and software are uploaded in the app stores every day; manual code inspection is unfeasible and automatizing the detection of malware and bugs becomes indispensable.
In this vibrant and evolving scenario, artificial intelligence (AI) is becoming the best ally for fighting the malicious user; they can handle a huge quantity of data and react quickly and autonomously to the environment changes. Unfortunately, we are not the only ones that can use AI support, but also the attackers can take advantage of this technology. Therefore, the research is moving towards designing sound code analysis techniques, which needs to be robust to adversarial attack.
- 23 March 2020 [14:00 - 15:00]
Speaker: Federica Baccini
Title: Network Analysis for the Integration of Histone Modification Data to Explain Haematopoiesis
Abstract: In this seminar a model that combines different techniques of network analysis is presented, together with its application to epigenetic data, in order to study haematopoiesis, i.e. the process of differentiation of blood cells.
Currently, there is not certainty about the phases of this hierarchical cell differentiation process, and existing hypotheses on its regulation in normal cells mostly rely on immunohistochemical studies. Until now, most of the epigenetic studies on haematopoiesis tend to concentrate on the individuation of differences in the regulation of gene expression between normal and diseased cells, rather than on the influence of epigenetic modifications on the normal process.
In this talk, the role of Epigenetics in the normal process of haematopoiesis will be explored by using similarity network fusion. Then, the classical hypothesis on the first hierarchical subdivision of haematopoiesis is tested by applying a greedy cut algorithm to networks of cellular types.
- 30 March 2020 [14:00 - 15:00]
Speaker: Francesco Odierna
Title: Federated Zero-Shot Learning: A Proposal
Abstract: Federated Learning is a machine learning setting in which a global model is trained using data stored on multiple remote devices. Zero-Short Learning allows to learn a classifier able to recognize new categories of instances without training examples. This presentation provides an overview of these two learning settings and some insights to merge them and obtain a new learning paradigm.
- 6 April 2020 [14:00 - 15:00]
Speaker: Eleonora Zedda
Title: Human-Robot Interaction in older adults with Mild Cognitive Impairment
Abstract: In the last decade, the aging of society is occurring worldwide. Aging has a considerable impact on
the health of the old people that suffer from cognitive impairments. The number of Mild Cognitive
Impairment (MCI) older adults is increasing. It becomes more and more important to support them
to avoid cognitive decline. Human-robot interaction can play an important role for this purpose.
This presentation provides an overview of the possibilities of using robots as assistive technologies
for older adults with cognitive impairments.
- 20 April 2020 [14:00 - 15:00]
Speaker: Giulia Punzi
Title: Maximal Common Subsequence Enumeration: How Graph Structure Helped Solve a String Problem
Abstract: Problems related to strings arise in almost every computer science field of research, as strings are one of the most common representations of data.
Graphs, on the other hand, are ubiquitous and intuitive data structures, used to represent relationships between concepts.
Graph algorithms are as essential and widespread as string ones.
In modern day scenarios, the need to combine string and graph algorithms is arising more and more often, as most information is textual,
while still interdependent in a linked way. We are thus led to believe that employing the underlying graph structure can help us design efficient algorithms to solve difficult string problems. In this talk, we will explore one instance of this phenomenon: the problem of enumerating Maximal Common Subsequences.
- 27 April 2020 [14:00 - 15:00]
Speaker: Francesco Landolfi
Title: A PyTorch-Geometric Tutorial
Abstract: This tutorial introduces PyTorch-Geometric, a PyTorch extension for managing and processing structured data such as graphs, 3D meshes, and point clouds.
Although PyTorch-Geometric main focus is on Machine Learning,
I will also present some common practices that can allow users to develop highly parallelizable algorithms for non-Euclidean data by exploiting SIMD architectures.
- 4 May 2020 [14:00 - 15:00]
Speaker: Alessio Molinari
Title: A critical reassessment of the SLD algorithm
Abstract: The talk will be about a critical re-examination of what we call the SLD algorithm, a well-known algorithm that, given a machine-learned classifier and a set of unlabelled data for which the classifier has output posterior probabilities and class prior probabilities, updates them both in an iterative, mutually recursive way, with the goal of making both more accurate. Since its publication in 2002, SLD has become the standard for improving the quality of posterior probabilities. However, since its application in my master thesis we have been wondering: is it really as effective as we thought it was?
- 11 May 2020 [14:00 - 15:00]
Speaker: Alessandro Bocci
Title: Securing Faas in the Fog
Abstract: Fog Computing is nowadays an established paradigm to bring the computational resources of the Cloud near the edge of the network
mainly to improve the Quality of Service of Internet of Things applications.
Recently FaaS (Function as a Service) has emerged as a paradigm based on functions that are executed in the Cloud in a serverless
When FaaS meets Fog, applications QoS improves, but new challenges arise, and security is one of those.
- 18 May 2020 [14:00 - 15:00]
Speaker: Asma Sattar
Title: Meta Level Hybrid Recommender System Algorithms: How Graphs can be used to improve performance of recommender systems
Abstract: Recommender systems use machine learning and data mining techniques to filter unseen information and predict whether a user would like a particular item. One research challenge in this field is to make useful recommendation from available set of millions of items with sparse ratings. A large number of approaches have been proposed aiming to increase accuracy, but they have ignored potential problems, such as sparsity and cold start problems. We have designed a novel hybrid recommendation framework that combines content-based filtering with collaborative filtering that overcome aforementioned problems.
Another research challenge is to deal with contextual information. Can we do something better with contextual information? How heterogeneous graphs can be used to represent recommender systems? Lets meet and discuss future of graph based recommender systems?
- 25 May 2020 [14:00 - 15:00]
Speaker: Daniela Rotelli
Title: Computational modeling of reading
Abstract: A better understanding of how children read and comprehend a short text, and what makes this process occasionally difficult, slow and inefficient is key to improving people's education level, professional qualification and social cohesion. By using a tablet device with a touchscreen for detecting the text focused by the child while reading, we enable for an automated and integrated modality for evaluating the decoding skills, the comprehension level and the reading speed. ReadLet intends to use and validate portable ICT technology and cloud computing to collect, time-align, integrate and analyse large streams of multimodal child reading data, elicited through ecological protocols at school.