GAN Dissection: Visualizing and Understanding Generative Adversarial Networks

By David Bau, Jun-Yan Zhu, Hendrik Strobelt, Bolei Zhou, Josh Tenenbaum, William T. Freeman, Antonio Torralba. 2018

513 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition, Generative Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Generative Models

Citations in the corpus

A Style-Based Generator Architecture for Generative Adversarial Networks (2018, background): “Yet the generators continue to operate as black boxes, and despite recent efforts [2], the understanding of various aspects of the image synthesis process, e.”
Regionwise Generative Adversarial Image Inpainting for Large Missing Areas (2019, background): “Some work [53], [54] tried to analyze the reason behind, but still cannot totally overcome this problem.”
Semantic Image Synthesis With Spatially-Adaptive Normalization (2019, background): “For example, class-conditional models [4,5, 34, 36, 38] leam to synthesize images given category labels.”
Rewriting a Deep Generative Model (2020, methodology): “But why is rewriting a deep generative model useful? A generative model enforces many rules and relationships within the generated images [7,36].”
Causal Interpretability for Machine Learning - Problems, Methods and Evaluation (2020, background): “Traditional Interpretability Interpretable Models: [106], [103], [64], [102], [65], [105], [50], [13], [40] Causal Interpretability Model-based: [77], [38], [12], [8], [112], [81], [7], Example-based: [35], [45], [39], [101], [76], [86], [73], [61], [60], [33], [34] Post-hoc: [47], [89], [66], [93],”
GAN Memory with No Forgetting (2020, background): “, low-level filters (those close to observations) are generally applicable, while high-level ones are task-specific [90, 93, 44, 4, 5, 20, 58, 98], the above discovery reveals an orthogonal dimensionality for transfer learning.”
Alias-Free Generative Adversarial Networks (2021, background): “While several ways of controlling the generative process have been found [8, 28, 10, 41, 24, 2, 7, 46, 6], the foundations of the synthesis process remain only partially understood.”
Self-Supervised Learning: Generative or Contrastive (2020, background): “GAN dissection [10] is a pioneer work in applying causal analysis into understading GAN.”
Paint by Word (2021, background): “One approach is to ﬁnd GAN la-tents to generate images that match a user’s painted intention: this can be done to match color [57] or a ﬁnite vocabulary of visual concepts [5, 4, 32] at a given location.”
GANSpace: Discovering Interpretable GAN Controls (2020, methodology): “A few methods provide useful control over spatial layout of the generated image [14, 26, 4, 3], provided a user is willing to paint label or edge maps.”
StyleSpace Analysis: Disentangled Controls for StyleGAN Image Generation (2020, methodology): “[5, 3] utilized semantic segmentation to analyze Progressive GAN [16] and detect causal units that control the presence of certain objects through ablation.”
Image2StyleGAN++: How to Edit the Embedded Images? (2019, methodology): “investigated what a GAN can and cannot generate by locating and manipulating relevant neurons in the activation tensors [4, 5].”
Inductive biases for deep learning of higher-level cognition (2020, background): “Ns), which are independent by construction, generally do not have a semantic interpretation (as a word or short phrase), whereas many units in slightly lower layers do have a semantic interpretation (Bau et al., 2018). Why not represent the causal structure with a directed graphical model? In these ”
On the "steerability" of generative adversarial networks (2019, background): “Deep learning for content creation The recent progress in generative models has enabled interesting applications for content creation [4, 13], including variants that enable end users to control and fine-tune the generated output [22, 26, 3].”