[Paper Review] On the Units of GANs (Extended Abstract)

David Bau, Jun-Yan Zhu+5 more|arXiv (Cornell University)|Jan 29, 2019

Cell Image Analysis Techniques10 references11 citations

TL;DR

This paper introduces a method to interpret and manipulate internal representations in Progressive GANs by identifying interpretable units that causally influence semantic objects like doors, windows, and chairs. By ablating or inserting these units, the authors demonstrate causal control over object generation, artifact reduction, and context-aware object placement, revealing that GANs learn structured, disentangled representations of scene concepts.

ABSTRACT

Generative Adversarial Networks (GANs) have achieved impressive results for many real-world applications. As an active research topic, many GAN variants have emerged with improvements in sample quality and training stability. However, visualization and understanding of GANs is largely missing. How does a GAN represent our visual world internally? What causes the artifacts in GAN results? How do architectural choices affect GAN learning? Answering such questions could enable us to develop new insights and better models. In this work, we present an analytic framework to visualize and understand GANs at the unit-, object-, and scene-level. We first identify a group of interpretable units that are closely related to concepts with a segmentation-based network dissection method. We quantify the causal effect of interpretable units by measuring the ability of interventions to control objects in the output. Finally, we examine the contextual relationship between these units and their surrounding by inserting the discovered concepts into new images. We show several practical applications enabled by our framework, from comparing internal representations across different layers, models, and datasets, to improving GANs by locating and removing artifact-causing units, to interactively manipulating objects in the scene. We will open source our interactive tools to help researchers and practitioners better understand their models.

Motivation & Objective

• To understand whether GANs learn internal representations corresponding to human-perceptible objects like doors, buildings, and trees.
• To develop a method for identifying and manipulating specific units within a GAN that causally affect the presence or absence of semantic objects.
• To debug GANs by locating and ablating units responsible for visual artifacts.
• To study contextual relationships between objects in GAN-generated scenes through controlled interventions.
• To enable interactive, object-level image manipulation using causal unit control.

Proposed method

• Units in feature maps are identified by upscaling and thresholding their activations, then measuring overlap with ground-truth semantic segmentation masks using IoU and pixel accuracy.
• A unit is classified as a class predictor if it achieves >0.75 pixel accuracy and >0.05 IoU with a semantic class after upscaling and thresholding.
• Causal effects are studied by ablating (setting to zero) or inserting (activating) specific units in the latent representation and observing changes in generated images.
• The impact of interventions is traced through layers by measuring normalized L1 changes in downstream feature maps, visualized as heatmaps and line plots.
• The method compares representations across layers, GAN variants, and datasets to analyze how training innovations affect interpretability.
• Ablation of 20 artifact-causing units is evaluated using FID and human preference scores to quantify quality improvement.

Experimental results

Research questions

• RQ1 Do specific units in a GAN's internal representation correspond to human-perceptible objects such as doors, windows, or chairs?
• RQ2 Can manipulating individual units cause the appearance or disappearance of specific objects in generated images?
• RQ3 How do contextual constraints in scenes affect the success of object insertion or removal via unit manipulation?
• RQ4 What is the causal relationship between internal units and visual artifacts in GAN outputs?
• RQ5 How do feature activations propagate through layers, and what determines whether an intervention results in a visible change in the final output?

Key findings

• Ablating 20 artifact-causing units reduced FID from 52.87 to 32.11 and increased human preference from 50.8% to 79.0%, demonstrating significant visual quality improvement.
• Units matching semantic classes (e.g., doors, windows) emerge prominently in layers 4–7 of a Progressive GAN, with minimal semantic content in earlier layers.
• The number of interpretable units increases with model quality, and innovations like batch normalization and pixelwise normalization enhance interpretability.
• Object removal is context-dependent: people and windows are easily erased in conference rooms, while tables and chairs in the same setting are harder to fully remove.
• Inserting door units in layer 4 causes doors to appear only in contextually appropriate locations (e.g., buildings), with attributes like size, color, and style matching the scene.
• Downstream layer analysis shows that interventions in layer 4 have amplified effects up to layer 12, with stronger signal propagation in contextually valid regions (e.g., buildings) than in invalid ones (e.g., sky or trees).