[Paper Review] Disentangling Factors of Variations Using Few Labels

Francesco Locatello, Michael Tschannen+4 more|arXiv (Cornell University)|Apr 30, 2020

Digital Media Forensic Detection62 references52 citations

TL;DR

This paper investigates how minimal, imprecise labeling of factors of variation (0.01–0.5% of data) enables reliable disentanglement in representation learning. Using a large-scale study of 52,000 models, it demonstrates that even weak supervision allows effective model selection and training, enabling state-of-the-art disentangled representations with minimal human labeling.

ABSTRACT

Learning disentangled representations is considered a cornerstone problem in representation learning. Recently, Locatello et al. (2019) demonstrated that unsupervised disentanglement learning without inductive biases is theoretically impossible and that existing inductive biases and unsupervised methods do not allow to consistently learn disentangled representations. However, in many practical settings, one might have access to a limited amount of supervision, for example through manual labeling of (some) factors of variation in a few training examples. In this paper, we investigate the impact of such supervision on state-of-the-art disentanglement methods and perform a large scale study, training over 52000 models under well-defined and reproducible experimental conditions. We observe that a small number of labeled examples (0.01--0.5% of the data set), with potentially imprecise and incomplete labels, is sufficient to perform model selection on state-of-the-art unsupervised models. Further, we investigate the benefit of incorporating supervision into the training process. Overall, we empirically validate that with little and imprecise supervision it is possible to reliably learn disentangled representations.

Motivation & Objective

• To investigate whether minimal supervision can enable reliable disentanglement in representation learning.
• To evaluate the impact of few, potentially imprecise labels on model selection and training for disentangled representations.
• To empirically validate the effectiveness of weak supervision in overcoming theoretical limitations of unsupervised disentanglement.
• To provide a reproducible benchmark for assessing disentanglement under limited labeling.

Proposed method

• A large-scale experimental setup trained over 52,000 models under well-defined, reproducible conditions.
• Leveraged existing state-of-the-art unsupervised disentanglement models as baselines for comparison.
• Applied sparse labeling (0.01–0.5% of data) with potentially incomplete or imprecise labels to guide model selection and training.
• Used labeled examples to select the best-performing unsupervised model from a pool of candidates.
• Integrated supervision directly into training to improve disentanglement performance.
• Evaluated disentanglement quality using standard metrics under controlled, reproducible settings.

Experimental results

Research questions

• RQ1 Can a small number of labeled examples (0.01–0.5%) enable effective selection of the best disentangled representation model?
• RQ2 Does incorporating weak supervision into training significantly improve disentanglement performance?
• RQ3 How robust are disentanglement methods to imprecise or incomplete labels in limited supervision?
• RQ4 Can supervision mitigate the theoretical limitations of unsupervised disentanglement learning?

Key findings

• A small number of labeled examples (0.01–0.5% of the dataset) is sufficient to reliably select the best-performing unsupervised disentanglement model.
• Even with imprecise and incomplete labels, supervision enables consistent improvement in disentanglement performance.
• Model selection based on few labels outperforms purely unsupervised approaches in terms of disentanglement quality.
• Incorporating supervision into training leads to more robust and disentangled representations than unsupervised training alone.