Tezos

Deep Neural Paintings

Creator

tz1WXed6b3wEPLjeQgxDqS8F7R7D8FHUarYh

Description

Deep Neural Paintings (2015–2017) - Triptych as shown at Ars Electronica (2017) Deep Neural Paintings interrogates how contemporary artificial intelligences see and how visual tradition—art, camouflage, and tactics of deception—can be reimagined in the age of deep learning. Framed as a 21st-century take on Cubism and WWI’s Dazzle, the series treats deep neural networks (DNNs) as both audience and co-author: it crafts encoded optical-illusions that are largely meaningless to humans yet confidently recognized by state-of-the-art image classifiers. Aim & Approach  The project translates the disruptive logic of historical dazzle—geometric, high-contrast patterns that confuse human perceptual judgment—into patterns and textures that disrupt machine perception. Using evolutionary algorithms and frameworks (e.g. Sferes) and adversarial methods, the work generates images and surface treatments that reliably fool discriminative models (e.g. CNNs trained on ImageNet), producing classifications with implausible certainty while remaining visually ambiguous to people. These encodings can be applied to two- and three-dimensional forms, suggesting practical and speculative interventions in both everyday objects and military contexts. Security, Ethics & Hacktivism  Deep Neural Paintings stages a conversation about automated decision-making in modern conflict and security infrastructures. If visual metadata and machine learning drive lethal outcomes, what role can art play as critique, countermeasure, or protective strategy? The series asks hard questions about trust in algorithmic systems, the ethics of automated targeting, and whether design practices can intervene to reduce harm—raising hacktivist and humanitarian stakes alongside artistic inquiry. Demonstration & Findings  Works produced in the series demonstrate that adversarially encoded images—often unrecognizable or absurd to human viewers—can lead advanced classifiers to declare familiar objects with >99.99% confidence. This highlights a systemic vulnerability shared by many discriminative AI techniques and suggests novel forms of disruptive camouflage and obfuscation that translate historical visual tactics into computationally aware strategies. Lineage & Influence  Rooted in the histories of Cubism and Dazzle painting, the project connects early 20th-century experiments in fragmented perspective and geometric abstraction to contemporary software aesthetics and machine vision. It also dialogues with recent research in adversarial networks and the work of practitioners exploring the political and cultural dimensions of AI. Credits & Inspiration  Informed by the pioneering research of J. Clune, A. Nguyen and J. Yosinski—whose work on neural visualization and adversarial examples exposed key limits of machine perception—this project sits at the intersection of art and AI. De Wilde collaborated with the latter to develop the pioneering interactive AI artwork The Innovation Engine (2015). In sum, Deep Neural Paintings is the outcome of an interdisciplinary inquiry that repurposes formal strategies from art history to both interrogate and actively intervene in the visual regimes of twenty-first-century machine vision and contemporary art.