Visualidades computacionais e a imagem-rede: reapropriações do aprendizado de máquina para o estudo de imagens em plataformas online

• Ano de defesa: 2019
• Autor(a) principal: André Goes Mintz
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; FAFICH - FACULDADE DE FILOSOFIA E CIENCIAS HUMANAS; Programa de Pós-Graduação em Comunicação Social; UFMG
• Palavras-chave em Português: Imagem; Métodos digitais; STS; Aprendizado de máquina; Visão computacional; Comunicação; Máquinas
• Link de acesso: http://hdl.handle.net/1843/31956

This research seeks to contribute do the field of Digital Methods for studies in Communication and Media by focusing, specifically, the issue of computational analysis of images through machine learning techniques. Aiming to overcome operational methodological perspectives which turn to quantitative approaches, this study proposes to reflect upon how digital images are theoretically considered in those efforts and how computational methods conform particular modes of seeing. The theoretical framework is mainly derived from Science and Technology Studies (STS) and, especially, Actor-Network Theory (ANT). Through concepts from these domains, images are understood as sociotechnical inscriptions, in a condition of ontological multiplicity. An uncertainty is therefore assumed regarding the possibility of treating images individually and, in contrast, it is proposed to approach them as effects of distributed relational materialities. These are central aspects of the conceptual hypothesis of the image-network, which is proposed by this thesis. This formulation is articulated to an effort of describing the operation of machine learning image recognition techniques based on artificial neural networks. Considering the relations among these models and large training data sets harvested from the internet, as well as their infrastructuralization tendencies, it is considered that they are important components of the contemporary visual field, generating computational visualitites with strong participation in datafication and algorithmic mediation processes to which images are subjected in online platforms. The application of these techniques as methodological resources is proposed, therefore, as a critical repurposing which considers methods as integral parts of the objects under scrutiny. This approach is exercised in a case study focusing on images published on Twitter during a media event. In a confluence of the theoretical and methodological discussions, a methodological device named Atlas for Image-Networks is finally proposed. It seeks to afford conditions for heuristic navigational practices through the images, while also attempting to preserve the ontological multiplicity of their instantiations.