Introduction to Artificial Social Intelligence

Uploaded by AUTH

The tutorial is an introduction to the main aspects of Artificial Social Intelligence (Social AI), the AI domain aimed at making machines socially intelligent, i.e., capable to make sense of the social landscape in the same way as people do. The focus will be on the most specific aspects of the field with respect to the rest of AI, including interdisciplinary connection with Human Sciences (Psychology, Anthropology, Sociology, etc.), methodological aspects of data collection, approaches for human behavior sensing, evaluation metrics, experimental design and Deep Learning methodologies for analysis and synthesis of human behavior. The tutorial will make use of data that will be shared with the participants and will provide the immediate opportunity to perform experiments in some of the main areas of Social AI, from the detection of mental health issues to generation of artificial emotions with virtual agents.

Quantitatively Assessing Explainable AI for Deep Neural Networks – a Crash Course

Uploaded by AUTH

Deep Neural Networks (DNNs) are increasingly pervasive into society, especially in decision-making, in applications involving humans or in high-stake applications. This prompts the need for transparency, which is one of the cornerstones of the EU guidelines for Trustworthy AI. For DNNs, it is often unfeasible to attain human-understandable interpretations of the predictive dynamics, hence approximate post-hoc explanations are used instead. These range from feature importance, to generating counterfactual data, and identifying important concepts and training data points. The approximate nature of these tools, though, raises critical questions concerning the quality of the explanations: are these explanations really faithful to the models inner dynamics? How sensitive are they to variations in input or model? Are they really interpretable to humans? Normally, due to the absence in explanation ground truths, many works applying Explainable AI (XAI) tools either fail to evaluate their quality, or rely on anecdotal evaluations, e.g., with user studies, which often results in poor evaluations due to subjectivity or difficulty in defining what constitutes a “valid explanation” for a given task.

This tutorial aims at exploring the recent developments in the topic of quantitative assessment of XAI tools. The first part will be dedicated to introducing the main concepts and methods for XAI applied to DNNs. Next, the axes of evaluation of explanations will be described. Finally, the main metrics and experimental settings for a sound evaluation of XAI tools will be illustrated, providing insights into what constitutes a “good metric” for evaluating explanations. The tutorial will conclude with an outlook on recent trends in the field.

On the possible outputs of human brains and intelligent machines

Uploaded by AUTH

In this presentation I will argue that present-day intelligent machines, featuring the presently available modes of input and of input analysis procedures, cannot possibly produce identical outputs with those of human brains. The latter produce discriminant responses or publicly accessible behavior, more generally, on the one hand and subjective conscious experiences, on the other. The former can output only behavior. Specifically, experiences come in several varieties: sensory, affective, noetic. Except for the possibility that the brain contains genetically programmed correlates of some experiences (e.g. general ideas) all varieties of experience appear to derive from the sensory ones (i.e.visual, auditory, tactile, kinesthetic, gustatory, olfactory). The way that environmental inputs are transformed into neural correlates of sensations is not known and neither is the structure of such correlates. This luck of knowledge constitutes the main reason that the present mode of input processing on the part of intelligent machines cannot be modified so as to produce the artificial analogues of such neural correlates. Consequently, the output of intelligent machines, unlike that of the brain, is limited to publicly accessible behavioral responses.

Looking into the Future: Forecasting Quantities with Deep Learning

Uploaded by AUTH

This lecture will cover recent advances in methodologies to forecast quantities using deep
neural networks with applications to autonomous agents, video streaming and network
traffic forecasting. We first briefly introduce sequence prediction problems introducing
the main architectural choices, such as RNNs, LSTMs and Transformers. Then we will
delve into forecasting of agent motion in different settings, reporting on our recent
research in social trajectory forecasting with the use of memory augmented neural
networks. Finally, we will conclude with recent results on large models for time series
forecasting and their application to network traffic estimation.

Cyberphysical AI: Towards a (Robot) AI Scientist

Uploaded by AUTH

Toward A(G)I, whatever definition one chooses for AGI, the main effort is finding the best possible self-supervised, unsupervised, and supervised objectives that scale to the foundation level. Taking a step back, some broader questions arise naturally. Are these objectives teaching our machines how the world looks or how we humans look at the world? Would it make sense to design machines and robots that progressively learn to interact with the world “scientifically”: from turning pixels to symbols, to posing scientific hypotheses for how the world works, then learning better scientific representations (that allow for experimenting with the world), and in the end defining autonomous ways for experimenting with the world and validating the hypotheses? In this lecture, I will present my work and vision for “Robot AI Scientists” that learn autonomously about the world and how to interact “scientifically” using AI that bridges data-based learning with physical understanding and cause-and-effect experimentation.

Trustworthy Machine Learning in Multimodal AI Applications: Case Studies and Perspectives

Uploaded by AUTH

Machine learning and deep learning models are the main engines in many multimodal AI applications, which are characterized by the fusion of multiple modalities of data streams. In this lecture, we highlight the trust and robustness challenges of machine learning that arises from data fusion. To do so, we present several case studies demonstrating how multimodal applications exacerbate existing challenges of trustworthy and robust machine learning. In a first case study, we investigate the impact of fusion depth on the robustness of multi-modal machine learning models, observing that model architecture could impact robustness. In a second case study, we investigate the impact of fusion modality on the robustness of multi-modal machine learning models, observing that fusion models are only as robust as their most susceptible modality. In another case study, we explore the impact of weight quantization techniques on the robustness of multimodal models, observing the need for modality-based quantization schemes. Through these case studies, we hope to share some perspectives on the unique trust and security challenges that arise in AI machine learning models in typical multimodal applications and offer insights to fortify such systems in real-world scenarios.