Counterfactuals uncover the modular structure of deep generative models

By Michel Besserve, Rémy Sun, Bernhard Schölkopf. 2018

110 citations recorded by the source corpus.

Deep generative models can emulate the perceptual properties of complex image datasets, providing a latent representation of the data. However, manipulating such representation to perform meaningful and controllable transformations in the data space remains challenging without some form of supervision. While previous work has focused on exploiting statistical independence to disentangle latent factors, we argue that such requirement is too restrictive and propose instead a non-statistical framework that relies on counterfactual manipulations to uncover a modular structure of the network composed of disentangled groups of internal variables. Experiments with a variety of generative models trained on complex image datasets show the obtained modules can be used to design targeted interventions. This opens the way to applications such as computationally efficient style transfer and the automated assessment of robustness to contextual changes in pattern recognition systems.

Subfields: Cognitive Neuroscience, Computer Vision and Pattern Recognition, Safety, Robustness & Interpretability, Causal Inference & Discovery, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Cognitive Neuroscience

In Computer Vision and Pattern Recognition

In Safety, Robustness & Interpretability

In Causal Inference & Discovery

In Signal Processing

Citations in the corpus

Toward Causal Representation Learning (2021, background): “done by ensuring that they are invariant across problems, exhibit sparse changes to actions or that they can be independently intervened upon [22], [30], [217].”
Towards Causal Representation Learning (2021, background): “by ensuring that they are invariant across problems, exhibit sparse changes to actions, or that they can be independently intervened upon [221, 21, 29].”
A theory of independent mechanisms for extrapolation in generative models (2020, methodology): “2018) and its superiority to statistical independence has been emphasized (Besserve et al. 2020; Locatello et al. 2018).”
Causality for Machine Learning (2019, methodology): “…or mechanisms, we discuss the special case of intervening upon the latent noise variables.19 One way to intervene is to replace noise variables with the corresponding values computed from other input images, a procedure that has been referred to as hybridization by Besserve et al. (2018b).”
Robustly Disentangled Causal Mechanisms: Validating Deep Representations for Interventional Robustness (2018, background): “The above deﬁnition on its own is likewise a contribution to the deﬁned but unused notion of extrinsic disentanglement in Besserve et al. (2018).”
Causal Interpretability for Machine Learning - Problems, Methods and Evaluation (2020, background): “Traditional Interpretability Interpretable Models: [106], [103], [64], [102], [65], [105], [50], [13], [40] Causal Interpretability Model-based: [77], [38], [12], [8], [112], [81], [7], Example-based: [35], [45], [39], [101], [76], [86], [73], [61], [60], [33], [34] Post-hoc: [47], [89], [66], [93],”
Self-Supervised Learning: Generative or Contrastive (2020, background): “In [12], researchers show that in GAN the features of cock is disentangled with the features of background.”
Interventional Few-Shot Learning (2020, background): “Nowadays, there are theoretical [3, 8] and empirical [77, 71] evidences showing that disentangled semantic manifolds emerge during training deep networks.”
Structural Autoencoders Improve Representations for Generation and Transfer (2020, methodology): “Inspired by [24], we refer to it as hybrid sampling: The model keeps a small set of N ( = 128 ) latent vectors stored (much smaller than the training dataset).”
Using latent space regression to analyze and leverage compositionality in GANs (2021, background): “Related to our work, Besserve et al. (2018) estimates the modular structure of GANs by learning a causal model of latent representations, whereas we investigate the GAN’s compositional properties using image inversion.”
Counterfactual Generative Networks (2021, background): “Besserve et al. (2019) also leverage the idea of independent mechanisms to discover modularity in pre-trained generative models.”
CausalVAE: Disentangled Representation Learning via Neural Structural Causal Models (2020, background): “[1] adopts dependent latent factors in the model.”
Causal Intervention for Weakly-Supervised Semantic Segmentation (2020, background): “The purpose of causal inference [39, 43] is to empower models the ability to pursue the causal effect: we can remove the spurious bias [5], disentangle the desired model effects [6], and modularize reusable features that generalize well [35].”