Abstract
Saliency maps have been used as one possibility to interpret deep neural networks. This method estimates the relevance of each pixel in the image classification, with higher values representing pixels which contribute positively to classification.The goal of this study is to understand how the complexity of the network affects the interpretabilty of the saliency maps in classification tasks. To achieve that, we investigate how changes in the regularization affects the saliency maps produced, and their fidelity to the overall classification process of the network.The experimental setup consists in the calculation of the fidelity of five saliency map methods that were compare, applying them to models trained on the CIFAR-10 dataset, using different levels of weight decay on some or all the layers.Achieved results show that models with lower regularization are statistically (significance of 5%) more interpretable than the other models. Also, regularization applied only to the higher convolutional layers or fully-connected layers produce saliency maps with more fidelity. © 2020 IEEE.

Abstract
Researchers in a variety of fields are increasingly adopting crowdsourcing as a reliable instrument for performing tasks that are either complex for humans and computer algorithms. As a result, new forms of collective intelligence have emerged from the study of massive crowd-machine interactions in scientific work settings as a field for which there is no known theory or model able to explain how it really works. Such type of crowd work uses an open participation model that keeps the scientific activity (including datasets, methods, guidelines, and analysis results) widely available and mostly independent from institutions, which distinguishes crowd science from other crowd-assisted types of participation. In this paper, we build on the practical challenges of crowd-AI supported research and propose a conceptual framework for addressing the socio-technical aspects of crowd science from a CSCW viewpoint. Our study reinforces a manifested lack of systematic and empirical research of the symbiotic relation of AI with human computation and crowd computing in scientific endeavors.

Abstract

Abstract
The increasing interest in connecting small sensors to the internet has led to the development of operating systems able to support all network, graphical and server functionalities over diverse embedded hardware. Globaltronic, a company based in Portugal, has developed an embedded computer called WiiPiiDo, powered by an ARM SoC (System on a Chip), which is highly specialized for IoT applications. It features NB-IoT - LTE Cat NB1 (Narrow Band IoT) to ensure robust connection to the Internet even in harsh conditions, and allows rapid development of complete IoT solutions for end-users. The development of a Linux image that exposes all the potential of the hardware platform is a must and will provide extra value to it. To create such an operating system, we examined the Yocto Project, which is a Linux building environment that is gaining a growing community of users, particularly enterprises. Nevertheless, Yocto is not the only choice for the embedded platform developer community. In fact, Armbian, a Debian/Ubuntu based distribution, appears as an popular alternative for embedded development in ARM boards. In this paper we show the steps we took from testing the first boot on the WiiPiiDo board until the development of the supporting operating system, finalizing with performance tests. We present a comparison of the two build systems that were used and report the results from the performance tests. © 2020, Springer Nature Switzerland AG.

Abstract
Tailoring of the user experience to each individuals needs and preferences can lead to more accessible solutions. Adaptation has a key role on matching the systems characteristics with the user needs. This can be achieved with personalization (system-driven adaptation) or customization (user-driven adaptation). Personalization have good results on matching the user needs but raises concerns about privacy and how the information is retrieved. Customizations require that the user manually choose the preferences configuration. This article proposed two versions of a web wizard to elicit the accessibility preferences. One version was based on a quiz and the other one presented small interactive activities. The activities version was proposed to help reduce the burden of configuration by implicitly eliciting the user preferences through interactive small activities. The preliminary results with healthy participants suggest that both versions obtained a positive evaluation. However, there was no major difference between each wizard. The causes of these findings are discussed.

Abstract
The human hand is the body part most frequently injured in occupational accidents, accounting for one out of five emergency cases and often requiring surgery with subsequently long periods of rehabilitation. This paper proposes a Virtual Reality game to improve conventional physiotherapy in hand rehabilitation, focusing on resolving recurring limitations reported in most technological solutions to the problem, namely the limited diversity support of movements and exercises, complicated calibrations and exclusion of patients with open wounds or other disfigurements of the hand. The system was assessed by seven able-bodied participants using a semistructured interview targeting three evaluation categories: hardware usability, software usability and suggestions for improvement. A System Usability Score (SUS) of 84.3 and participants' disposition to play the game confirm the potential of both the conceptual and technological approaches taken for the improvement of hand rehabilitation therapy. © 2020 IEEE.