Spring of AI Seminar Series
16 October 2019 | 9:30 am | Sala Consiglio – Ground Floor, West Building
SPEAKER: Paolo Fiorini, University of Verona
TITLE: Embodied AI: towards autonomous functions in medicine, surgery and manufacturing
ABSTRACT: In the past 20 years, the robotics laboratory ALTAIR of the University of Verona has been carrying out research in applications of robotics to safety-critical situations, in particular, medicine and surgery. The focus has been on understanding the main constraints of the surgical process, from diagnosis to therapy, with the goal of increasing patient safety and comfort. It became evident that some form of intelligent supervision is necessary to ensure that information is properly exchanged and timed, and that actions are accurate and safe. Thus the concept of autonomous robot has emerged as the research paradigm that can integrate and demonstrate the technologies of embedded AI. In surgery, autonomy is explored with respect to the execution of surgical tasks and bedside assistance during robotic surgery, and in rehabilitation, we are developing an intelligent exoskeleton for the upper body. Furthermore, the learning and execution technologies developed for medicine will be soon applied to the manufacturing process, in the context of a new research laboratory in intelligent production, e.g. Industry 4.0. In this talk, I will give an overview of the background justifying the need of autonomous functions in safety-critical applications and then I will describe some of our current research in the development of the building blocks of a cognitive robot for surgical applications.
30 October 2019 | 2:30 pm | Sala Consiglio – Ground Floor, West Building
SPEAKER: Giuseppe Jurman, FBK
TILE: OmicsCNN: adding structural information to Convolutional Neural Networks
ABSTRACT: Convolutional Neural Networks (CNNs) is a popular deep learning architecture widely applied in different domains, in particular in classifying over images, for which the concept of convolution with a filter comes naturally. Unfortunately, the requirement of a distance (or, at least, of a neighbourhood function) in the input feature space has so far prevented its direct use on data types such as omics data. However, a number of omics data are metrizable, i.e., they can be endowed with a metric structure, enabling to adopt a convolutional based deep learning framework, e.g., for prediction. We propose a generalized solution for CNNs on omics data, implemented through a dedicated Keras/PyTorch layer. For
transcriptomics data, we combine Gene Ontology semantic similarity and gene co-expression to define a distance; the function is defined through a multilayer network where 3 layers are defined by the GO mutual semantic similarity while the fourth one by gene co-expression. For metagenomics data, a metric can be derived from the patristic distance on the phylogenetic tree, used together with a sparsified version of MultiDimensional Scaling to embed the phylogenetic tree in a Euclidean space. As a general tool, feature distance on omics data is enabled by the layer OmicsConv, obtaining OmicsCNN, a dedicated deep learning framework. In particular, OmicsConv is taking care of passing to the following convolutional
layer not only the data but also the ranked list of the neighbourhood of each sample, thus mimicking the case of image data, transparently to the user. Here we demonstrate OmicsCNN on gut microbiota sequencing data, for Inflammatory Bowel Disease (IBD) 16S data, first on synthetic data and then a metagenomics collection of gut microbiota of 222 IBD patients.
13 November 2019 | 2:30 pm | Sala Consiglio – Ground Floor, West Building
SPEAKER: Luciano Serafin, FBK
TITLE: Bayesian Inference in Hybrid Statistical Relational Learning
ABSTRACT: One of the most important foundational challenges of Statistical relational learning is the development of a uniform framework in which learning and logical reasoning are seamlessly integrated. The presentation focuses on a Bayesian approach SRL. In particular, we propose to model the knowledge in hybrid domains (i.e., domains that contains relational structure and continuous features) with a set T of FOL axioms and a set F of (in)dependence hypothesis that specifies a probability over the interpretations that satisfies the axioms in T and the independence hypothesis encoded in F. This probability is represented with a density function p(x | T,F) that is computed via variational inference. The presentation will introduce the theory and a simple example implemented in a system called SemCla.
27 November 2019 | 2:30 pm | Sala Consiglio – Ground Floor, West Building
SPEAKER: Marco Turchi, FBK
TITLE: Machine Translation for Machines
ABSTRACT: With the rapid growth of cloud-based software-as-a-service offerings, a variety of affordable high-performance Natural Language Processing (NLP) tools has become easily accessible. Unfortunately, however, the language coverage of most of these tools is still limited to English or, in the best cases, to a few other languages. For the vast majority of not covered languages, the easiest solution is to adopt a pivoting, translation-based, approach. Translating from one of these languages (say, Croatian) into a high-resource one (say, English) gives the possibility to i) use existing tools trained for the high-resource language to process the translated text, and ii) projecting their output back to the original text. In this pivoting framework, the standard quality criteria for machine translation (MT) might represent a sub-optimal objective. While MT traditionally aims at generating “fluent and adequate” output to be consumed by human speakers of the target language, here the goal is to produce translations that are “useful and easy to process” by a machine trained for the downstream NLP task at hand. The paradigm shift brought by this “machine-oriented” scenario, in which the machines are
the new target opens a novel perspective to MT research and raises interesting deep learning challenges. In this talk, I will introduce the problem, discuss a possible solution based on reinforcement learning and present our latest results in several tasks (including sentiment classification, hate-speech detection and sentence classification).
4 December 2019 | 2:30 pm | Sala Consiglio – Ground Floor, West Building
SPEAKER: Chiara Ghidini, FBK
TITLE: On the dimensions of predictive process monitoring
ABSTRACT: Predictive process monitoring aims at predicting the future of an ongoing process execution by learning from past historical business process executions. Different approaches have been proposed in the literature in order to provide predictions on the outcome, the remaining time, the required resources as well as the remaining activities of an ongoing execution, by leveraging information related to the control flow, the data flow, or even unstructured text contained in event logs, recording information about process executions. The approaches can be of different nature and, some of them also equipped to offer users support in tasks such as parameter selection. This talk will provide an overview of existing approaches, with particular emphasis on the work carried out in the PDI group in the last 5 years, as well as future challenging research directions in the field of predictive process monitoring.
18 December 2019 | 2:30 pm | Sala Consiglio – Ground Floor, West Building
SPEAKER: Andrea Passerini, UniTN
TITLE: Hybrid Relational Learning
ABSTRACT: tbd
Embodied AI: towards autonomous functions in medicine, surgery
and manufacturing