Synthesizing Programs for Images using Reinforced Adversarial Learning

Abstract

Advances in deep generative networks have led to impressive results in recent years. Nevertheless, such models can often waste their capacity on the minutiae of datasets, presumably due to weak inductive biases in their decoders. This is where graphics engines may come in handy since they abstract away low-level details and represent images as high-level programs. Current methods that combine deep learning and renderers are limited by hand-crafted likelihood or distance functions, a need for large amounts of supervision, or difficulties in scaling their inference algorithms to richer datasets. To mitigate these issues, we present SPIRAL, an adversarially trained agent that generates a program which is executed by a graphics engine to interpret and sample images. The goal of this agent is to fool a discriminator network that distinguishes between real and rendered data, trained with a distributed reinforcement learning setup without any supervision. A surprising finding is that using the discriminator's output as a reward signal is the key to allow the agent to make meaningful progress at matching the desired output rendering. To the best of our knowledge, this is the first demonstration of an end-to-end, unsupervised and adversarial inverse graphics agent on challenging real world (MNIST, Omniglot, CelebA) and synthetic 3D datasets.

12 pages, 13 figures

Cited by in corpus (45)

• A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications
• Abstraction and Analogy-Making in Artificial Intelligence
• Learning to Infer and Execute 3D Shape Programs
• Content and Style Aware Generation of Text-line Images for Handwriting Recognition
• Object Discovery with a Copy-Pasting GAN
• Computer-Aided Design as Language
• Learning to Paint With Model-based Deep Reinforcement Learning
• Pragmatic inference and visual abstraction enable contextual flexibility during visual communication
• Top-Down Synthesis for Library Learning
• Neural Painters: A learned differentiable constraint for generating brushstroke paintings
• Deep Reinforcement Learning in Computer Vision: A Comprehensive Survey
• Learning to Sketch with Deep Q Networks and Demonstrated Strokes
• Write, Execute, Assess: Program Synthesis with a REPL
• Exploiting Hierarchy for Learning and Transfer in KL-regularized RL
• Interactive Sketch & Fill: Multiclass Sketch-to-Image Translation
• Learning To Simulate
• From Scratch to Sketch: Deep Decoupled Hierarchical Reinforcement Learning for Robotic Sketching Agent
• Unsupervised Image to Sequence Translation with Canvas-Drawer Networks
• Who Left the Dogs Out? 3D Animal Reconstruction with Expectation Maximization in the Loop
• Differentiable Drawing and Sketching
• Modeling 3D Shapes by Reinforcement Learning
• RLSS: A Deep Reinforcement Learning Algorithm for Sequential Scene Generation
• Learning Differentiable Programs with Admissible Neural Heuristics
• Learning Task-General Representations with Generative Neuro-Symbolic Modeling
• Generating new concepts with hybrid neuro-symbolic models
• Unsupervised and interpretable scene discovery with Discrete-Attend-Infer-Repeat
• Adversarial Intrinsic Motivation for Reinforcement Learning
• Creative Procedural-Knowledge Extraction From Web Design Tutorials
• Generative Adversarial Networks are Special Cases of Artificial Curiosity (1990) and also Closely Related to Predictability Minimization (1991)
• Interactive Neural Style Transfer with Artists
• Stroke-based Character Reconstruction
• Structured agents for physical construction
• Combining Semantic Guidance and Deep Reinforcement Learning For Generating Human Level Paintings
• Learning Evolved Combinatorial Symbols with a Neuro-symbolic Generative Model
• Painting Many Pasts: Synthesizing Time Lapse Videos of Paintings
• Emergence of Writing Systems Through Multi-Agent Cooperation
• PLANS: Robust Program Learning from Neurally Inferred Specifications
• Generating Diverse Programs with Instruction Conditioned Reinforced Adversarial Learning
• SketchTransfer: A Challenging New Task for Exploring Detail-Invariance and the Abstractions Learned by Deep Networks
• Learning to Infer User Interface Attributes from Images
• Contrastive Active Inference
• Channel Decomposition into Painting Actions
• Unsupervised Program Synthesis for Images By Sampling Without Replacement
• When Autonomous Systems Meet Accuracy and Transferability through AI: A Survey
• Relational Mimic for Visual Adversarial Imitation Learning