Understanding and Improving Interpolation in Autoencoders via an Adversarial Regularizer

Abstract

Autoencoders provide a powerful framework for learning compressed representations by encoding all of the information needed to reconstruct a data point in a latent code. In some cases, autoencoders can "interpolate": By decoding the convex combination of the latent codes for two datapoints, the autoencoder can produce an output which semantically mixes characteristics from the datapoints. In this paper, we propose a regularization procedure which encourages interpolated outputs to appear more realistic by fooling a critic network which has been trained to recover the mixing coefficient from interpolated data. We then develop a simple benchmark task where we can quantitatively measure the extent to which various autoencoders can interpolate and show that our regularizer dramatically improves interpolation in this setting. We also demonstrate empirically that our regularizer produces latent codes which are more effective on downstream tasks, suggesting a possible link between interpolation abilities and learning useful representations.

Cited by in corpus (19)

• Interpolation Consistency Training for Semi-Supervised Learning
• RelGAN: Multi-Domain Image-to-Image Translation via Relative Attributes
• Learning protein conformational space by enforcing physics with convolutions and latent interpolations
• A Comprehensive Overview and Survey of Recent Advances in Meta-Learning
• Deep hybrid model with satellite imagery: how to combine demand modeling and computer vision for behavior analysis?
• Interpolation-based semi-supervised learning for object detection
• Fair Mixup: Fairness via Interpolation
• Towards Domain-Agnostic Contrastive Learning
• Adversarial Domain Adaptation with Domain Mixup
• APRICOT: A Dataset of Physical Adversarial Attacks on Object Detection
• Few-Shot Unsupervised Continual Learning through Meta-Examples
• Unsupervised Meta-Learning through Latent-Space Interpolation in Generative Models
• Embracing the Dark Knowledge: Domain Generalization Using Regularized Knowledge Distillation
• Generative Model without Prior Distribution Matching
• Enabling the Network to Surf the Internet
• Divergent Search for Few-Shot Image Classification
• Semantics-Guided Representation Learning with Applications to Visual Synthesis
• Multi-class Novelty Detection Using Mix-up Technique
• TRIP: Refining Image-to-Image Translation via Rival Preferences