Keynote Lectures

Forecast the Forecasting
Marco Cristani, University of Verona, Italy, Italy

A Review on Malicious Facial Image Processing and Possible Counter-Measures
Jean-Luc Dugelay, EURECOM, France, France

Combining Images and Words in Deep Networks that Identify People from Body Shape
Alice Joan O'Toole, The University of Texas at Dallas, United States, United States

Letting Go of the Numbers: Measuring AI Trustworthiness
Carol J. Smith, Carnegie Mellon University, Software Engineering Institute, United States, United States

Abstract
Traditionally, the topic of forecasting belongs to the area of probability and statistics, and mainly concerns numerical series in the econometric and financial fields. However, several prediction-related topics are emerging in the field of pattern recognition, such as predicting trajectories, human poses, behaviors, and many others. Some of them are not just limited to academic research, having important implications in the industrial field. This talk offers an updated overview of this situation, showing some of the most effective technologies, and introducing little-explored but very promising applications where traditional forecasting and machine learning could, and should, meet.

Abstract
Recent advances in visual sensor technologies and computer vision enable to create more and more efficient face recognition systems, and more generally speaking, facial image processing tools. Such new performances generate some ethical and societal concerns. In parallel, such advances, in deep learning for example, contribute to the proliferation of malicious digital manipulations. The main objectives for designing such attacks is to access resources illegally, to harm some individuals or to make ineffective some technologies. In this presentation, we give an overview of existing attacks and related techniques attached to some applications in the domains of biometrics (e.g. identity spoofing), of medias (e.g. fake news), of video surveillance (e.g. de-anonymization). Some recent works investigate potential counter-measures for detecting such attacks.

Abstract
Many important applications of person identification occur at distances and viewpoints from which the face is not visible or insufficiently resolved to be useful. In these cases, body shape can serve as a biometric modality that can generalize across distance and viewpoint variation. I will discuss our approach to body identification, which combines standard object classification networks with representations based on word-based descriptions of bodies. In our previous work, we showed that a relatively small number of linguistic descriptors (e.g., broad shoulders, muscular, curvy, pear-shaped, short-legs) can be used in conjunction with a PCA-based model of 3D bodies to generate an accurate 3D reconstruction of a body. We reasoned that this description might also be sufficient for identifying individual bodies. I will overview our work implementing body shape identification networks with and without linguistic training. We compared these models on their ability to identify people from body shape using images captured across a large range of distances/views (close-range, 100m-1000m, and in images from unmanned aerial vehicles [UAV]). Accuracy, as measured by identity-match ranking and false accept errors, was surprisingly good. Non-linguistic models were generally more accurate for close-range images, whereas linguistic models fared better at farther distances and for UAV. Fusion of the linguistic and non-linguistic embeddings improved performance in nearly all cases. This work shows that linguistic and non-linguistic representations of body shape can offer complementary identity information about bodies that can improve identification in applications of interest. At a broader perspective, the evolution of language to be a highly effective and concise form of communication may underlie its current success in a wide range of machine learning applications.

Abstract
AI systems need to be designed to work with, and for, the people using them. A person’s willingness to trust a particular system is based on their expectations of the system’s behavior. Their trust is complex, transient, and personal – it cannot easily be measured. However, an AI system’s trustworthiness can be measured. A trustworthy AI system demonstrates that it will fulfill its promise by providing evidence that it is dependable in the context of use, and the end user has awareness of its capabilities during use. We can measure reliability and instrument systems to monitor usage (or lack thereof) quantitatively. However, AI’s potential is bound to perceptions of its trustworthiness, which requires qualitative measures to fully ascertain. Doing AI well requires a reset – letting go of (some of) the numbers, and learning new methods that provide a more complete assessment of the system.