Courses

Speech-to-speech translation (STST) can be generally seen as the combination of three sub-tasks: (i) transcribing speech to text in a
source language (ASR), (ii) translating text from a source to a target language (MT) and (iii) generating speech from text in a target
language (TTS).  Significant progress has been recently made in these three distinct tasks as well as in their joint combination.
Remarkably the state of the art in both pipelined and end-to-end STST systems is achieved by deep learning models that have
fundamental characterists in common: they are all sequence-to-sequence models (seq2seq) models with an encoder, a decoder and an attention network.
This course will start with a general introduction to deep learning, and will then focus on sequence-to-sequence models for MT, ASR and
TTS, including variations and combinations of them

Conversational Agents are one of the most impressive evidence of the recent resurgence of Artificial Intelligence. In fact, there is now a high expectation for a new generation of dialogue systems that are able to naturally interact and assist humans in a number of scenarios, including virtual coaches, personal assistants and automatic help desks.
This course introduces some of the main technological challenges behind the development of conversational agents, under the perspective of Computational Linguistics. We will address Human-to-Human dialogue phenomena, as well as methodologies to collect Human-to-Machine dialogues. We will introduce the characteristics of the components of a dialogue system, including the slot filling component, the dialogue manager and the generation component. Finally, we provide an overview of commercial tools for building chatbots as well as of evaluation methodologies for assessing the performance of a dialogue system

Security is a core requirement for manufacturers, developers, service providers and other stakeholders who produce and use IoT devices to develop systems and applications for Smart Cities and/or IoT systems, like those for industry 4.0 or connected cars. Securing these is a major challenge, and failure to do so can result in significant harm to individuals and businesses. The course will address the challenges to securing IoT and Smart Cities devices used to deploy systems and applications.
The course will focus on a hardware-led approach to create stronger security for IoT and Smart Cities systems, once software-based approaches are not sufficient to guarantee the application or system security. The course proposes three general areas of guidance:
Addressing fundamental controls for securing IoT devices. Developers must address two basic questions related to IoT and Smart Cities systems: “What are we trying to protect?” and “What is required to enable protection?” The answers result in a shortlist of fundamental controls necessary to implement hardware-based security. The core requirement is a trusted IoT operating environment enabled with a secure boot process that is impervious to attack. This requires a root of trust forged in hardware, which establishes a chain of trust for all IoT subsystems.
Using a Security by Separation approach. Security by Separation is a classic, time-tested approach to protecting computer systems and the data contained therein. Separation means functions cannot see or access other functions without authorization. By separating and restricting the availability and use of assets, security is enforced according to prescribed policy. In this way, software that implements one function does not have to trust software, which is implementing another function – each is separated from each other. The course focuses on IoT and Smart Cities systems that can retain their security attributes even when connected to open networks. It hinges on the use of logical separation created by a variety of methods: the most secure is hardware-based virtualization, which entails systems used to simulate, isolate and control IT assets. There are pros and cons to using other separation methods, but they can serve as interim implementations, such as paravirtualization, hybrid virtualization and Linux containers.
Enforcing secure development and testing. Finally, developers must provide an infrastructure that enables secure debug for IoT and Smart Cities systems during product development and testing. Normally, hardware debugging through JTAG allows the user to see the entire system. A secure system needs to maintain the separation of assets even when using hardware debugging. By embracing these initial areas of focus, designers can take action to create secure operating environments for IoT and Smart Cities systems by means of secure application programming interfaces (APIs). The APIs will create the glue to enable secure inter-process communications between disparate software and applications. The course will focus in IoT and Smart Cities systems to illustrate these challenges.

1) Review of Bayes decision theory. Bayes theorem, optimality of Bayes decision rule, rewriting of the discriminant functions into equivalent forms, case studies.

2) Review of artificial neural networks (ANN). Definitions, MLPs and deep architectures, supervised learning, mixtures of experts, autoencoders, application to pattern recognition, universality, estimation of class-posterior probabilities, estimation of scaled-likelihoods, radial basis functions.

3) Parametric estimation techniques. Equivalence between supervised and unsupervised setups. Maximum likelihood (ML) approach. Mixture densities and GMMs. From GMMs to k-Means clustering to competitive neural nets.

4) Nonparametric estimation. General framework. Parzen window, kn-nearest neighbor. Examples. Pros and cons.

5) Density estimation via Parzen neural networks (PNN). Training algorithm. Practical matters, model selection via cross-validated likelihood, application to pattern classification. Overview of the theoretical properties of PNNs (complexity, modeling capabilities, asymptotic convergence in probability). Graphical demos. Application to sex determination from human crania.

6) Nonparametric pdf estimates via soft-constrained ANNs that satisfy Kolmogorov's axioms of probability. Markov Chain Monte Carlo solution to the computation of the numeric integral of the ANN; technique for sampling from the ANN. Results of simulations.

7) Neural Mixture Models (NMM) for the estimation of mixture densities. The relevance of mixture densities, difference between mixture densities and mixture density models, ML soft-constrained training of the NMM. Results of simulations.

8) Sequence processing: hard-constrained RBF-based ML density estimation over sequences of patterns encoded via the Echo State Network. Application to emotion recognition from speech signals.

9) Graph processing: hard-constrained RBF-based ML density estimation over graphs (i.e., structured  patterns, or relations) encoded via the recursive/graph neural networks. Applications to density estimation over graphs, graph clustering, and graph classification.

New generation technologies and device miniaturization are boosting the development of wearable unobtrusive embedded systems, capable to acquire and process physiological signals, whereas making sense of the data retrieved with such devices requires real-time execution of advanced algorithms on resource-constrained platforms. Hence, they require a multi-modal approach to reach miniaturized form factors at extreme energy efficiency. In this course, we will give an overview of the HW-SW co-design of such systems, ranging from acquisition circuits to algorithms. We will describe embedded system design for wearable application in 2 classes of physiological signals: inertial data and biopotentials (EEG and EMG in particular). The course will describe basic concepts in signal acquisition from digital Inertial Measurement Units, analog/digital EMG and EEG. It will also explore advanced processing techniques and machine learning optimized for embedded platform to extract meaningful information close to the device and preserving energy efficiency.  Examples from rehabilitation, prosthetics and HMI domain will be presented

Network science has emerged as a branch of study focusing interest on the connectivity interactions between elements of a system. The most central object of study in network science is the so-called complex networks. Complex weblike structures describe a wide variety of systems of high technological and intellectual importance. For example, the cell is best described as a complex network of chemicals connected by chemical reactions; the Internet is a complex network of routers and computers linked by various physical or wireless links; fads and ideas spread on the social network, whose nodes are human beings and whose edges represent various social relationships; the World Wide Web is an enormous virtual network of Web pages connected by hyperlinks. These systems represent just a few of the many examples that have recently prompted the scientific community to investigate the relationship between the topology of complex networks and the dynamics that take place on them. A complex network is just a graph with several non-trivial topological properties, not present in simple models of networks. Some of them are scale-free degree distributions, high clustering coefficients (i.e. more triangles than expected in a random network), assortativity (correlations between connected nodes’ degrees), and community structure. On the contrary, simple graphs such as random networks or grids show a homogeneous structure in which all nodes are almost indistinguishable, unlike what is observed in real networks. In this course, we will review the state-of-the-art in network science and put the focus on the applications, and open problems faced so far.

Conversational Agents are one of the most impressive evidence of the recent resurgence of Artificial Intelligence. In fact, there is now a high expectation for a new generation of dialogue systems that are able to naturally interact and assist humans in a number of scenarios, including virtual coaches, personal assistants and automatic help desks. This course introduces some of the main technological challenges behind the development of conversational agents, under the perspective of Computational Linguistics. We will address Human-to-Human dialogue phenomena, as well as methodologies to collect Human-to-Machine dialogues. We will introduce the characteristics of the components of a dialogue system, including the slot filling component, the dialogue manager and the generation component. Finally, we provide an overview of commercial tools for building chatbots as well as of evaluation methodologies for assessing the performance of a dialogue system.

The course introduces the foundations and recent advancements of machine translation (MT), probably the most prolific application sector in computational linguistics. Machine translation deals with the automatic translation of speech or text between two languages. This technology is considered strategic for the integration of Europe as well as for the global market. Most internet companies (Google, Microsoft, Facebook, eBay, Amazon, Airbnb, ...) have their own MT research teams in order to support their geographic expansion. This course focuses on the statistical or machine learning approach which, after a long hegemony of phrase-based models, recently moved to neural network based models. While focusing on all relevant aspects related to phrase-based and neural MT, in particular modelling, training, decoding, we will also look at use cases of MT in the translation industry, ranging from e-commerce to the support of professional translators.

The goal of the course is to; provide basic concepts in Requirements Engineering (RE); learn and practice with state of the art RE methods and techniques; learn about some open problems in RE and the applicability of techniques such as NLP and automated reasoning techniques to propose novel solutions.