AI, Logistics and Complex Systems: perspectives from an AI user
Research Center for Advanced Science and Technology, The University of Tokyo (JAPAN)
AI is now being applied in a wide variety of fields. However, as its application has progressed, various issues have also come to light. I would like to introduce the possibilities and challenges of AI applications, especially in the fields of pedestrian flow and logistics, in which I have been deeply involved. We analyzed human flow to reduce congestion at the Tokyo Olympics last year, and have recently been using AI to monitor social distance in public places. However, the detection accuracy is not yet sufficient due to the variety of environments in the real field. In logistics, unmanned operations are now fairly common, and AI is playing an active role in efficient warehouse operations and vehicles’ route planning. But it has yet to provide a holistically optimal solution. In my presentation, I would like to discuss how AI should be utilized in the future considering these current conditions.
SESSION 1 - SHADES OF MACHINE LEARNING: CAUSAL, LIFELONG AND RELATIONAL
Department of Computer Science, University of Pisa (ITALY)
Graphs are an effective representation for complex information, providing a straightforward means to bridge numerical data and symbolic relationships. The talk will provide an easy paced introduction to the lively field of deep learning for graphs, its applications and open research challenges. Dealing with graph data requires learning models capable of adapting to structured samples of varying size and topology, capturing the relevant structural patterns to perform predictive and explorative tasks while maintaining the efficiency and scalability necessary to process large scale networks. The talk will first introduce foundational aspects and seminal models for learning with graph structured data. Then it will discuss the most recent advancements in terms of deep learning for network and graph data, including learning structure embeddings, graph convolutions, attentional models and graph generation.
Introduction to Continual Learning
Department of Computer Science, University of Pisa (ITALY)
Learning continually from non-stationary data streams is a long-standing goal and a challenging problem in machine learning research. Naively fine-tuning prediction models only on the newly available data often incurs in Catastrophic Forgetting or Interference: a sudden erase of all the previously acquired knowledge. On the other hand, re-training prediction models from scratch on the accumulated data is not only inefficient but possibly unsustainable in the long-term and where fast, frequent model updates are necessary. In this lecture we will discuss recent progress and trends in making machines learn continually through architectural, regularization and replay approaches. We identify Continual Learning as a promising approach and key technological enabler towards the design of systems compliant with the Sustainable AI principles. Then, we present Avalanche, an open-source end-to-end library for continual learning based on PyTorch and discuss possible real-world applications.
Causality-inspired ML: what can causality do for ML? The domain adaptation case
University of Amsterdam and MIT-IBM Watson AI Lab
Applying machine learning to real-world cases often requires methods that are robust w.r.t. heterogeneity, missing not at random or corrupt data, selection bias, non i.i.d. data etc. and that can generalize across different domains. Moreover, many tasks are inherently trying to answer causal questions and gather actionable insights, a task for which correlations are usually not enough. Several of these issues are addressed in the rich causal inference literature. On the other hand, often classical causal inference methods require either a complete knowledge of a causal graph or enough experimental data (interventions) to estimate it accurately. Recently, a new line of research has focused on causality-inspired machine learning, i.e. on the application ideas from causal inference to machine learning methods without necessarily knowing or even trying to estimate the complete causal graph. In this talk, I will present an example of this line of research in the unsupervised domain adaptation case, in which we have labelled data in a set of source domains and unlabelled data in a target domain ("zero-shot"), for which we want to predict the labels. In particular, given certain assumptions, our approach is able to select a set of provably "stable" features (a separating set), for which the generalization error can be bound, even in case of arbitrarily large distribution shifts. As opposed to other works, it also exploits the information in the unlabelled target data, allowing for some unseen shifts w.r.t. to the source domains. While using ideas from causal inference, our method never aims at reconstructing the causal graph or even the Markov equivalence class, showing that causal inference ideas can help machine learning even in this more relaxed setting.
SESSION 2 - AI FOR HUMAN MACHINE INTERACTION
Wearable devices and sensors technologies
Department of Informatics, Systems and Communication, University of Milano-Bicocca (ITALY)
The latest advances in wearable devices and sensing technologies have enriched the potential of human-machine interaction (HMI), enabling the development of new applications especially in real-life environments. Artificial intelligence plays a significant role within new scenarios that suffer from noise, are strongly human-centered and often require severe constraints in terms of computational costs and real-time processing. The AI4HMI session aims to provide an overview of emerging technologies and their challenges. It offers concrete examples of wearable sensing applications, from BCI to VR, and understanding of behavior using physiological signals. students will have the opportunity to test some wearable devices, experimenting with various issues related to the definition of a real experimental protocol that allows data acquisition, in a tech-demo session.
Expanding Human Capacities with Wearable AI
Faculty of Industrial Design Engineering, TU Delft (THE NETHERLANDS)
Evolution has always been the main driving force behind slow but steady change for the human brain and body. However, living in the Information era, our perceptual and cognitive capacities cannot simply rely on natural evolution to keep up with the immense advancements in emerging technologies. On one hand, technologies we use daily (PCs, smartphones, wearables, etc.) remain largely uninformed about our perceptual levels, cognitive states and physical needs, thus forming an “awareness gap” between the human (user) and the machine (system). On the other hand, the same technologies, if properly actuated, can be prominent drivers for human augmentation: improved capture and sensing (e.g., cameras and sensors), in-situ information presentation (e.g., AR/VR displays and ubiquitous projection), and technologies for implicit input and adaptive control (e.g., eye-tracking, electromyography (EMG), and electroencephalography (EEG)). The impact of human augmentation can be substantial, multifaceted, and scalable, benefiting the able-bodied and the impaired, the layman and the expert, the institutional and the commercial, the individual and the masses. This lecture will focus on research conducted at the crossroads of Human-Computer Interaction (HCI), Ubiquitous Computing, AI, Cognitive Psychology, and Neuroscience. The highlighted prototypes include AI-powered wearables that augment perceptual and cognitive capacities, from employing sensor fusion to hosting Conversational Agents. The showcased research entails the digitalisation, introduction, and evaluation of previously in-lab clinical methods into the users’ natural settings. Here, technology is deemed the medium for attaining the sought-after impact in the wild, such as improving memory recall, increasing productivity levels, extending situational awareness, etc.
Sensors at Work: Personal Devices for Supporting Well-Being and Productivity
Faculty of Informatics,
University of Italian Switzerland - USI Lugano (SWITZERLAND)
Personal devices – such as smartphones, smartwatches, and the like – are nowadays pervasively used and can support their users in executing a plethora of tasks, including, e.g., tracking water intake or monitoring sleep quality. To this end, the devices leverage the sensing capabilities hidden within the devices, e.g., to monitor heart rate or physical activities. The increasing accuracy and diversity of sensors available on personal devices, along with advances in automated analysis of personal data, is pushing the potential of such sensor-based, personal informatics systems even further. In this lecture, we will provide an overview of both commercially available and research-grade systems for monitoring human behavior using personal devices. We will then explain the fundamental sensing primitives that allow such systems to work, outlining the related challenges and opportunities, with a specific attention on systems can be used to support people’s well-being and productivity at work.
Learning from Biological Intelligence of Insects
Research Center of Advanced Science Science and Technology, The University of Tokyo (JAPAN)
To elucidate the dynamic information processing in a sensor and a brain underlying adaptive behavior obtained through evolution (i.e., biological intelligence), it is necessary to understand the behavior and corresponding neural activities. This requires animals which have clear relationships between behavior and corresponding neural activities. Insects are precisely such animals and one of the adaptive behaviors of insects is high-accuracy odor source orientation, which is not yet available in conventional approaches.
To examine the neural basis of the odor-source orientation behavior, we have employed a strategy that tackles the question at multiple levels from genes, single cells of the neural system to the actual behavior. We have developed an insect-robot hybrid system, which moves depending on the behavioral or the neural output of a brain, as a novel experimental system. The robot is controlled by the behavior of an insect tethered on the robot or by the neural activity of the insect brain. This system has contributed to better understanding of the behavioral and neural basis of adaptive behavior. We also have developed highly sensitive olfactory sensors based on olfactory receptor proteins of insects using a genetic engineering.
At first in this lecture, strategy of odor navigation of a male silkmoth and its neural basis revealed by using multidisciplinary approaches will be shown. Second, the extent of adaptation in the behavioral strategy, as governed by the neural system and investigated via a robotic implementation, will be introduced. An odor-source searching drone with insect antenna as olfactory sensor will also be shown.
Our multidisciplinary research will enable us to use the full potential of the features of insect sensors and brains as model systems for understanding the dynamical sensory and neural substrates of adaptive behaviors (i.e., biological intelligence) and its application.
How affective Virtual Reality can contribute to AI research
Department of Mechanical Engineering,
Politecnico di Milano (ITALY)
The interest in human emotions is constantly growing. There is an interest in designing and creating tools capable of detecting the emotion that the person is experiencing in a given context at a given moment. On the other hand, there is an interest in how to generate certain emotions. Virtual reality (VR) is one of the most interesting technologies for generating emotions. In VR, designers can play with multiple senses and can create imaginary worlds where everything can happen. In this way, it is possible to generate emotions of any kind. Such environments can also be used to study human reactions, and they can also be used to improve algorithms able to detect human emotions. In this sense, VR can be used to generate data and train artificial intelligence algorithms.
Department of Informatics, Systems and Communication, University of Milano-Bicocca (ITALY)
During this session, students will have the opportunity to test some wearable devices, experimenting with various issues related to the definition of a real experimental protocol that allows data acquisition.
AI AND PHYSICS
Open-source software development in quantum computing
Quantum computing brings the opportunity to devise AI algorithms based on a new model of information processing, a new kind of computer architecture and a new kind of processing units' characteristics. Research and development efforts are underway to design so-called quantum hardware and quantum software. The ecosystem that has emerged in quantum computing comprises tools for quantum machine learning and machine learning applied to quantum computing and physics research. In this talk I'll review the state of the art of quantum computing, from the perspective of the data scientist or AI researcher approaching the open-source software stack. I will provide a brief overview of of the quantum computing stack and actively developed projects. In the second part of the lecture, I will focus on best practices for modern software development, with hands-on examples on version control, unit testing, documentation and release, using Python as an example.
A quantum solution to a classic Machine Learning problem
Department of Physics “Giuseppe Occhialini”, University of Milano-Bicocca (ITALY)
In this lecture we will solve a classic Machine Learning task, MNIST classification, leveraging quantum layers. We will then compare and discuss the two approaches, classical and quantum.
Ultrafast Machine Learning Inference at the Large Hadron Collider
Institute for Particle Physics and Astrophysics,
ETH, Zürich (SWITZERLAND)
At the CERN Large Hadron Collider, protons are brought to collide hundreds of millions of times per second. The collision debris allows us to study the fundamental building blocks of the universe and look for hints of new forces and particles. The vast majority of the collision data are immediately discarded by a real-time event filtering system due to storage and computational limitations. While most of these data are uninterresting, signals of new physics might be inadvertendly thrown away in the process. The first stage of this event filtering system consists of hundreds of field-programmable gate arrays (FPGAs), tasked with rejecting over 98% of the proton collisions within a few microseconds. With the start of High Luminosity LHC in 2029, a more granular detector and more particles per collision will increase the event complexity significantly, and ultimately require the FPGA farm to process an amount of data comparable to 5% of the total internet traffic. In this talk, I will discuss how we are using real-time Machine Learning on FPGAs to process and filter this enormous amount of data in the pursuit of new physics. I will discuss methods and tools for designing low-latency, efficient algorithms that run on FPGA hardware and, finally, I will explore how real-time anomaly detection can be used to process and collect data in a way never before performed at colliders.
Graph Neural Networks for Particle Physics Department of Physics “Giuseppe Occhialini”, University of Milano-Bicocca (ITALY) and INFN (Section of Milano-Bicocca)
The study of the collision events at the CERN Large Hadron Collider with Machine Learning algorithms presents unique challenges. Apart from the need to process very large data samples in a small amount of time, the algorithms need to be adapted to the specific structure of the dataset, consisting on the collection of all the particles produced in a given collision event. This object collection is unordered and heterogeneous (different types of particles), and there are strong event-by-event quantum fluctuations that change the total number of particles produced and the composition of the event. On the bright side, the collection of particles obeys certain fundamental rules, that can be exploited to simplify the processing. First of all, the system is permutation invariant: given any specific listing of the particles, the exchange of any two particle positions will lead to the same physical system. Second, most of the relevant information in the event can be attributed to “local regions” in the event, e.g. from a subgroup of particles produced in the decay of a single heavier particle. This “locality”, however, needs to be defined in a complicated and generally unknown latent space. A Graph Neural Network (GNN) is a type of neural network that matches particularly well the previous conditions, and is hence a very good candidate to study large systems of particles. In this talk, I will introduce the general structure of a GNN, and describe how these networks are being used in recently proposed algorithms to perform important event-reconstruction tasks at the LHCb and CMS experiments at CERN.
4 - AI, Language and Computation
Topic Modelling Meets Neural Language Model
Department of Informatics, systems and Communication, University of Milano-Bicocca (ITALY)
Topic models are statistical methods that aim to extract the hidden topics underlying a collection of documents. The main challenge related to topic modelling relates to the definition of training mechanisms that could capture the semantics of the document collection, at a reduced computational complexity. In this talk, we will initially present traditional statistical topic models - such as Latent Dirichlet Allocation – to then focus on the necessity of modeling the semantics of the language through their extensionswith Neural Language Models. We will also address the problem of model settings, evaluation and interpretability, highlighting the main open problems and research directions.
A Gentle Introduction to Bayesian Optimization and Its Application to Hyperparameter Tuning
Department of Economics, Quantitative Methods and Enterprises Strategies, University of Milano-Bicocca (ITALY)
Bayesian optimization is a sample efficient sequential method for the global optimization of a black-box, expensive, multiextremal, and possibly noisy function. A real-life example of this class of problems is the optimal tuning of a Machine Learning algorithm’s hyperparameters, the main application field considered in this talk. The basic idea, in Bayesian Optimization, is to combine learning-&-optimization: a probabilistic surrogate model of the function to be optimized is trained on the observations sequentially collected along the trials, and an acquisition function is used to choose the next trial – or a bunch of trials – to evaluate. The acquisition function deals with the exploration-exploitation dilemma depending on the surrogate’s predictive mean and standard deviation, with many possible alternatives offering different trade-off mechanisms. Different options are also available for the probabilistic surrogate model: Gaussian Process (GP) regression is best suited for the optimization over continuous search spaces while other approaches, such as Random Forests or GPs with ah-hoc kernels, deal with complex search spaces spanned by nominal, numeric and conditional variables. This talk offers an introduction to these topics, with a focus on the application of Bayesian Optimization for Hyperparameter Tuning. Moreover, it also provides a discussion on tools, other real-life applications, and recent advances, such as Bayesian Optimization under unknown constraints, leading to Automated Machine Learning on tiny devices (AutoTinyML), Bayesian Optimization with multiple information sources, leading to Green Machine Learning (GreenML), and Bayesian Optimization under multi/many-objectives leading to Fair-&-Green Machine Learning. Finally, an overview about open challenges and perspectives will be provided.
Profiling Neural Language Models
Institute for Computational Linguistics "Antonio Zampolli"
, CNR, Pisa (ITALY)
The field of Natural Language Processing (NLP) has seen an unprecedented progress in the last years. Much of this progress is due to the replacement of traditional systems with newer and more powerful algorithms based on neural networks and deep learning. This improvement, however, comes at the cost of interpretability, since deep neural models offer little transparency about their inner workings and their abilities. Therefore, in the last few years, an increasingly large body of work has been devoted to the analysis and interpretation of these models.
This presentation will be divided into two parts. In the first part, we will briefly introduce state-of-the-art Neural Language Models (NLMs) and discuss their characteristics. In the second part we will cover the most commonly applied analysis methods for understanding the inner behavior of NLMs based on Transformer architectures and how they implicitly encode linguistic knowledge.
Why is Imagination Mightier than Knowledge? Draw feature concepts with your pens, mightier than swords Department of Systems Innovation, The University of Tokyo (JAPAN)
A feature concept is a conceptual illustration of the abstract information about expected concept, knowledge, scenario, or structural outline of the target system to be acquired from data. For example, trees and clusters are feature concepts for decision tree learning and clustering respectively. In this talk, feature concepts are introduced as the essence of the data-federative innovation process. Since before the speaker gave the name "feature concepts" (FC), they have been elicited so far in or via creative communication among stakeholders in places of science and markets of data when they talked about what they like to know and what are the data they need. Recently, in our academia-industry collaboration laboratory on Data Federative Innovation Literacy (DFIL), participants started to draw FCs with reinforcing, fostering, and eliciting their imagination about the future societies they would find. With feature concepts, they explain data and tools they need. Without feature concepts, they often just desire data and seek knowledge about AI, without understanding about the contribution of mathematics to the embedded algorithms. By using FCs, we will learn at least a basic level of Einstein's famous words "imagination is more important than knowledge" - he may have been drawing FCs to live beyond data and beyond our present knowledge about AI.
Living Labs contribute to enhancing participant's Organizational Citizenship Behavior (OCB)
Department of Architecture, MIE University (JAPAN)
Without communication about peoples' real lives and the thoughts behind the data, we cannot realise the essential value of the data and the intention of its use. The Living Labs, where participants from various positions come together to share issues, perspectives and wisdom to come up with solutions, might be a place to actualise this. However, no one seems to have grasped the whole picture of Living Labs. Therefore, we have organised the history of the global development of Living Labs, the types, scope, methods and examples of activities, and arrived at the hypothesis that Living Labs have the effect of enhancing the organisational citizenship behaviour of participants. How does this elevation of 'organisationalcitizenship behaviour' actually come about? And what effects does it have? We would like to share examples that may substantiate the author's hypothesis while also opening up the participants' interest and contributing to the further development of the discussion.
SESSION 5 - AI AND HEALTH
Argument-Based Computational Persuasion for Behaviour Change in Healthcare
Free University of Bozen-Bolzano (ITALY)
Automated persuasion systems (APS) aim to persuade a user to believe something by entering into a dialogue in which arguments and counterarguments are exchanged. To maximize the probability that an APS is successful in persuading a user, it can identify a global policy that will allow it to select the best arguments it presents at each stage of the dialogue whatever arguments the user presents. However, in real applications, such as for healthcare, it is unlikely the utility of the outcome of the dialogue will be the same, or the exact opposite, for the APS and user. In order to deal with this situation, games in extended form have been harnessed for argumentation in Bi-party Decision Theory. This opens new problems that we address in this lecture: (1) How can we use Machine Learning (ML) methods to predict utility functions for different subpopulations of users? and (2) How can we identify for a new user the best utility function from amongst those that we have learned? To this extent, we develop two ML methods, EAI and EDS, that leverage information coming from the users to predict their utilities. EAI is restricted to a fixed amount of information, whereas EDS can choose the information that best detects the subpopulations of a user. We evaluate EAI and EDS in a simulation setting and in a realistic casestudy concerning healthy eating habits. Results are promising in both cases, but EDS is more effective at predicting useful utility functions.
Deep Learning in medical imaging
Department of Mathematics and Computer Science , University of Calabria (ITALY)
In recent years, Deep Learning techniques have been widely used in several healthcare applications, such as in the support to medical imaging diagnostic and computer-assisted surgery via detection, segmentation, and classification of specific pathologies or elements of medical interest. Such approaches emerged as powerful tools to improve healthcare assistance and pave the way to effective personalized medicine and treatment information. In this talk, we will provide an overview of the most used Deep Learning approaches for the analysis of pre- and intra-operative images. Then, we will illustrate recent advancements, innovative solutions, and real-world applications proposed in the field, with a special focus on the automatic instance segmentation task.
Bayesian (Causal) networks for Healthcare, Medicine and Biology
Department of Informatics, Systems and Communication, University of Milano-Bicocca (ITALY)
Health care and medicine were one of the first areas where artificial intelligence was applied, although initially with little impact on health care itself. Most of the impact of early AI in Medicine (AIME) research was in terms of the development of new AI methods. With the increasing availability of health-care data, there is now renewed interest in AIME, however, this time with the promise of having impact on healthcare. In this lecture we introduce Bayesian networks, causal networks and structural causal models and show how they are used to address and solve different problems in healthcare and medicine. The lecture starts by giving basic definitions and introduces paradoxes warning us about what, at the current state of the art, can be achieved and can not be achieved when only data are available. In particular, the ladder of causation of Pearl is described and explained. Then, formal definitions of Bayesian networks, causal network and structural causal models are given together with the concept of counterfactual which should be the basis for clinical decision making. The lecture closes by showing how causal networks can be learnt from observational and interventional from different data sources and by describing Bayesian network models for treatment of chronic kidney disease and prognosis of cardiovascular outcomes.
SESSION 6 - AI AND SUSTAINABILITY
The AI Carbon Footprint and Responsibilities of AI Scientists
Department of Electronic Engineering and Information Technology, University of Naples "Federico II"
This lecture examines ethical implications of the growing AI carbon footprint, focusing on the fair distribution of prospective responsibilities among groups of involved actors. First, major groups of involved actors are identified, including AI scientists, AI industry, and AI infrastructure providers, from datacenters to electrical energy suppliers. Second, responsibilities of AI scientists concerning climate warming mitigation actions are disentangled from responsibilities of other involved actors. Third, to implement these responsibilities nudging interventions are suggested, leveraging on AI competitive games which would prize research combining better system accuracy with greater computational and energy efficiency. Finally, in addition to the AI carbon footprint, it is argued that another ethical issue with a genuinely global dimension is now emerging in the AI ethics agenda. This issue concerns the threats that AI-powered cyberweapons pose to the digital command, control, and communication infrastructure of nuclear weapons systems.
Adding intelligence to ageing and longevity: a call to action for all of us National Innovation Center for Ageing, Newcastle (UK)
Intelligence is everywhere – driven by human genius and hunger for progress, evolving and advancing as we push the boundaries of what is possible. Daily new discoveries and inventions demonstrate the extraordinary brilliance of people to create tools and solutions unimaginable only a few years earlier. Intelligence is in data – from research, from a myriad of initiatives and projects, current, past and future. But Intelligence is also in human experience, shared knowledge and wisdom. And it’s found in emotions – there are deep insights in what people and their loved ones communicate, feel, want and need. Crucial yet disparate and largely imperceptible fragments of knowledge, collected over the phases of our lifetime. Imagine if we integrated these different sources intelligence – and had the power to harness people’s lived experience, their professional experience, amplify greater understanding of people’s aspirations and desires, and merge the strengths and capabilities of different generations and cultures? We aim to combine these valuable assets and knowledge across humans and leverage artificial intelligence and big data, with deliberation, ethics and technical ingenuity. We call it Ageing Intelligence®.
Computer Science and Sustainability: actual challenges and future perspective and the contribution of Informatics Europe
Department of Computer, Control, and Management Engineering "Antonio Ruberti", University of Rome "La Sapienza" (ITALY)
Digitalization of society demands response across all sectors, industries, and communities: not only to advance work on how to avoid societal disruptions, but to examine our own practices and their potential impact on the fairness and justice of our future. Concerns around sustainability form part of responsible innovation (RI) approaches, which are an important element in shaping the future of R&D across Europe. Informatics Europe represents the public and private research community of informatics in Europe and neighboring countries. Bringing together university departments, research laboratories and industry, it creates a strong common voice to safeguard and shape quality research and education in informatics in Europe. Representing over 150 members across Europe, Informatics Europe is gathering efforts for sustainability in informatics and through informatics. At our European Computer Science Summit in 2021, we brought together all levels of researchers, public policymakers and industry representatives across and beyond Europe to discuss sustainability with the main theme “Informatics for a Sustainable Future”. Recently, Informatics Europe have initiated a Sustainability Working Group. The main goal of this workgroup is to create a network for the diffusion of knowledge on the topic of Sustainable Computer Science and the application of Computer Science to Sustainability. The actual challanges and future perspectives for Sustainable Computer Science and the application of Computer Science to Sustainability are here analized and explained.
Faithing multidimensional poverty through AI
Department of Informatics,
Systems and Communication, University of Milano-Bicocca (ITALY)
Poverty is a multidimensional concept: the focus on financial resources alone does not capture people’s needs and quality of life. Being poor means in fact also a lack of access to resources enabling a minimum standard of living and participation in the society. The AMPEL (Artificial intelligence facing multidimensional poverty) project considers elderly people and consider data not only on income and wealth, but also on material and social deprivation that are rarely collected or known by public welfare institutions, making it difficult to intercept those who require more support. A dataset of heterogeneous data is used to extract the indicators needed to define a poverty risk. Unsupervised models are considered to unveil hidden patterns that can identify variational factors as separate sources of influence that can be used to characterize different levels of risk. Then supervised approaches are adopted for building predictive models, and hierarchical approaches are investigated for analyzing data focusing on different views as, for example, gender or size of municipalities.