30 September 2020

Due to the current situation this event will be postponed to September/October 20202 Recent striking success in Artificial Intelligence have made the public believe that in a not so far future machines could be even more intelligent than human beings. The actual and possible developments of Artificial Intelligence open up a series of striking, deep and pressing questions such as: – Can a computer ever think in the way a human being does? – Can a computer have a mind and conscious experiences, such as thoughts, desires, and emotions? – What is artificial intelligence? Is it the same as human intelligence? Are they even comparable or are they something essentially different? – Can a machine be morally responsible for its actions? Can a machine be good or evil? What other moral considerations are related to AI? With the goal of enhancing their scientific and educational collaborations around those important topics, the Swiss AI Lab IDSIA USI-SUPSI and the USI Master in Philosophy Program are organising in Lugano on May 29-30 2020 an international meeting on current trends and perspectives in the Philosophy of AI.
Auditorium USI

13 February 2020 - 13 February 2020

Probabilistic sentential decision diagrams are a class of probabilistic graphical models natively embedding logical constraints within a “deep” layered structure with statistical parameters. They thence induce a joint probability distribution over the involved Boolean variables that sharply assigns probability zero to states inconsistent with the logical constraints. In this presentation, I will first introduce and motivate such probabilistic circuits. I will then present a set-valued generalisation of the probabilistic quantification in these models, that allows to replace the sharp specification of the local probabilities with linear constraints over them, In doing so, a (convex) set of joint probability mass functions, all consistent with the assigned logical constraints, is induced.
Manno, Galleria 1, 2nd floor, room G1-201 @12:00

16 January 2020 - 16 January 2020

Most literature in the Philosophy of Computing stresses the dual, abstract and physical nature of computational systems. Under many respects, this debate reduces to the problem of explaining the relation which has traditionally been expressed in terms of the duality between specification and implementation. When this problem is analysed from the point of view of the notion of information though, computational systems require to be described at several levels of abstraction, and at each an appropriate notion of information is required. With such a conceptual tool in place, correctness of computational artifacts is adequately defined at functional, procedural and executional levels. A correct physical computational system is one which satisfies all such layers. This tripartite notion of correctness based on information is in turn essential to provide the basic elements of an appropriate logical analysis of efficiency, correctness, explanation and resilience for computational systems.
Manno, Galleria 1, 2nd floor, room G1-201 @12:00

18 November 2019

Cancer is a disease characterized by the accumulation of alterations to the genome, which selectively make cancer cells fitter to survive, proliferate and move. The understanding of progression and evolution models that underlie this processes, i.e., the characterization of sequences of alterations that lead to the emergence of the disease, is a topic attracting much attention. Of course, the problem of reconstructing such models is not new; in fact, several methods for inferring progression models (and phylogenies) from cross-sectional samples have been developed since the late 90s. Recently, we have proposed a number of algorithms to reconstruct cancer clonal evolutionary models from a variety of cross-sectional data types, either "ensemble", "bulk" or "single cell". We perform our reconstructions using a variety of algorithms based on a “probability raising” score that guarantees statistical dependencies on the inferred precedence relations. Our methods are complementary to traditional phylogeny reconstructions ones. Within this context, we have proven the correctness of our algorithms and characterized their performance. Our algorithms are collected in a R BioConductor package “TRanslational ONCOlogy” (TRONCO) that we have successfully used as part of our "Pipeline for Cancer Inference" (PiCnIc) to analyse Colorectal Cancer (CRC) data from TCGA. The newest addendum to TRONCO is the “Temporal oRder of Individual Tumors” (TRaIT) a new collection of algorithms that can be used for single-cell (and multi-region) progression analisys of cancers.
Room G1-201, Galleria 1, @12h00

14 November 2019

In this talk, we will discuss two different types of optical sensors based on luminescence: one is used for oxygen sensing type sensors, and the other is an «optical nose» for fingerprinting of substances. The problem of these types of sensors is that the measured signal is influenced by many components (like mirrors, lasers, electronics). Unfortunately these dependencies cannot be modeled mathematically in a simple way. So typically, complicated and empirical mathematical models are used, which are then fine tuned for each sensor in what is called calibration. But do we need to model those effects? Or there is another way? We will describe how the use neural networks can dramatically change how to build and use these sensors, without the need for any complicated mathematical model. Introducing neural networks in optical sensors typically does not require deep networks. However, there are several aspects that are very different from classical neural networks models which will be discussed here: one example overall is overfitting, that in this case requires completely different approaches to be dealt with. We will bring at the talk a prototype of a portable, low-cost sensor that we are currently developing at TOELT to demonstrate how a low-cost sensor can be built and used.
Manno, Galleria 1, 2nd floor, room G1-201 @12:00

30 October 2019

When building Bayesian networks with the help of domain experts, often properties of monotonicity arise, such as veterinarians expressing that “under more severe conditions, seeing the more severe symptoms becomes more likely”. It is well known that human decision makers will not use a network in their daily practice if such common properties of monotonicity are clearly violated, not even if the network shows overall high performance. In this talk, I will introduce some properties of monotonicity for Bayesian networks and further focus on the different roles that monotonicity has in engineering networks, both in building them by hand, in learning them from data and in fine-tuning.
Manno, Galleria 1, 2nd floor, room G1-201 @12:00

22 October 2019

The field of Geometric Packing problems has attracted the attention of many researchers in the late years. Generally speaking, in this setting we are given a region in the two-dimensional plane and a set of rectangles and the goal is to pack a subset of them inside the given region in such a way that they do not overlap and some given objective function is optimized. In this talk we will review our recent developments on two of these problems in the framework of Approximation Algorithms: Strip Packing and Geometric Knapsack. In the first problem, the goal is to pack all the rectangles into a strip of fixed width so as to minimize the final height of the packing, while in the second one the goal is to pack a subset of the rectangles of maximum profit into a rectangular region of fixed size. We will also discuss applications and open questions regarding the mentioned problems, the talk is meant to be accessible for non-experts.
Manno, Galleria 1, 2nd floor, room G1-204 @12h00

16 October 2019

Glyco-bioinformatics is an emerging subfield of bioinformatics aimed at expediting research in the field of glycomics. Unfortunately, the development of sugar-based virtual structures is made difficult by some structural features of sugar such as their high charge density, conformational flexibility and the torsional angles between glycosidic bonds. As a consequence, automated prediction of the binding poses of long sugar with proteins (that is a pivotal aspect of many biological processes) has been evaded so far, also due to the solvation/desolvation, weak surface complementarity and large electrostatic interactions of sugar/protein interactions. My PhD activity is aimed at overcoming these limits. To this aim, I have implemented a new computational method based on incremental docking that has been so far successfully applied to two important biological interactions such as that of heparin with the HIV-1 p17 matrix protein in the field of AIDS and VEGF with its VEGF receptor-2 in the field of tumor neovascularization. Perspective developments include the development of an algorithm able to automatize the developed computational methods and their application to other sugar/protein interactions of biological importance.
Manno, Galleria 1, 2nd floor, room G1-201 @12:00

The recent advances in Deep Learning made many tasks in Computer Vision much easier to tackle. However, working with a small amount of data, and highly imbalanced real-world datasets can still be very challenging. In this talk, I will present two of my recent projects, where modelling and training occur under those circumstances. Firstly, I will introduce a novel 3D UNet-like model for fast volumetric segmentation of lung cancer nodules in Computed Tomography (CT) imagery. This model highly relied on kernel factorisation and other architectural improvements to reduce the number of parameters and computational load, allowing its successful use in production. Secondly, I will discuss the use of representation learning or similarity metric learning for few-shot classification tasks, and more specifically its use in a competition at NeurIPS 2019 and Kaggle. This competition aimed to detect the effects of over 1000 different genetic treatments to 4 types of human cells, and published a dataset composed of 6-channel fluorescent microscopy images with only a handful of samples per target class.
Manno, Galleria 1, 2nd floor @12h00