[Paper Review] Bias and Generalization in Deep Generative Models: An Empirical Study

Shengjia Zhao, Hongyu Ren+4 more|arXiv (Cornell University)|Nov 8, 2018

Generative Adversarial Networks and Image Synthesis53 citations

TL;DR

The paper introduces a cognitive-psychology inspired framework to empirically study bias and generalization in deep generative models by probing low-dimensional features and examining single and multi-feature generalization patterns across GANs and VAEs.

ABSTRACT

In high dimensional settings, density estimation algorithms rely crucially on their inductive bias. Despite recent empirical success, the inductive bias of deep generative models is not well understood. In this paper we propose a framework to systematically investigate bias and generalization in deep generative models of images. Inspired by experimental methods from cognitive psychology, we probe each learning algorithm with carefully designed training datasets to characterize when and how existing models generate novel attributes and their combinations. We identify similarities to human psychology and verify that these patterns are consistent across commonly used models and architectures.

Motivation & Objective

• Characterize inductive bias and generalization behavior of deep generative models in high-dimensional image spaces.
• Develop a framework inspired by cognitive psychology to probe models via low-dimensional, interpretable features.
• Evaluate consistency of observed generalization patterns across architectures (GANs and VAEs), datasets (e.g., CLEVR, MNIST), and hyperparameters.

Proposed method

• Project high-dimensional image space onto interpretable low-dimensional feature spaces (e.g., numerosity, color, size, location).
• Train GANs and VAEs on controlled datasets and measure the induced feature distributions p(z) versus learned distributions q(z).
• Use single-feature probing to study impulse response and convolution-like generalization patterns.
• Examine multi-feature generalization via precision-recall on the support of z and assess memorization vs. novel combination generation.
• Assess independence of features by varying other features and observing stability of z-generalization.
• Visualize and compare generalization behaviors across model types and architectures.

Experimental results

Research questions

• RQ1 How do deep generative models generalize when trained on data with a single value (impulse response) for a feature?
• RQ2 How does multi-modal training data affect the model's tendency to memorize versus generate novel feature combinations?
• RQ3 Do single-feature generalization patterns (e.g., convolution, prototype enhancement) hold across GANs and VAEs and across datasets?
• RQ4 Are feature generalizations independent, such that analysis of one feature can be considered separately from others?
• RQ5 How does the model preserve marginal distributions while generating novel combinations across multiple features?

Key findings

• Generative models do not reproduce the exact training feature values; they display biased distributions, often centered near the training value with over-/under-shoot tendencies.
• When training data has multiple modes for a feature, the model behavior resembles a linear filter (convolution with an impulse response) except when modes are close, where prototype enhancement occurs.
• For color proportions, the learned distribution is approximately Gaussian centered at the trained value, with sharper variance for extreme proportions, consistent with Weber’s law.
• With multiple features, models memorize small training sets but generate novel combinations as the number of combinations increases, while preserving individual feature marginals.
• Recall is high (all training combinations appear in q(z)) but precision decreases as the number of combinations grows, indicating generation of novel combinations.
• Independence across features is observed; changing other features does not dramatically alter the learned distribution for a given feature, though higher variance can promote mode merging.