A Disentangling Invertible Interpretation Network for Explaining Latent Representations

By Patrick Esser, Robin Rombach, Björn Ommer. 2020

91 citations recorded by the source corpus.

Neural networks have greatly boosted performance in computer vision by learning powerful representations of input data. The drawback of end-to-end training for maximal overall performance are black-box models whose hidden representations are lacking interpretability: Since distributed coding is optimal for latent layers to improve their robustness, attributing meaning to parts of a hidden feature vector or to individual neurons is hindered. We formulate interpretation as a translation of hidden representations onto semantic concepts that are comprehensible to the user. The mapping between both domains has to be bijective so that semantic modifications in the target domain correctly alter the original representation. The proposed invertible interpretation network can be transparently applied on top of existing architectures with no need to modify or retrain them. Consequently, we translate an original representation to an equivalent yet interpretable one and backwards without affecting the expressiveness and performance of the original. The invertible interpretation network disentangles the hidden representation into separate, semantically meaningful concepts. Moreover, we present an efficient approach to define semantic concepts by only sketching two images and also an unsupervised strategy. Experimental evaluation demonstrates the wide applicability to interpretation of existing classification and image generation networks as well as to semantically guided image manipulation.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Representation & Self-Supervised Learning

How researchers see this paper

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

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Representation & Self-Supervised Learning

Citations in the corpus

Taming Transformers for High-Resolution Image Synthesis (2020, methodology): “[17, 68] demonstrate similar gains when using an unconditional normalizing flow for the second stage, and [55, 56] when using a conditional normalizing flow.”
High Fidelity Visualization of What Your Self-Supervised Representation Knows About (2021, methodology): “To allow for interpretable manipulation, a second invertible network (Esser et al., 2020) is trained using labels to disentangle the factors of variations present in the representation.”
GAN Inversion: A Survey (2021, methodology): “The IIN method [159] learns invertible disentangled interpretations of variational autoencoders (VAEs) [160].”
A comprehensive taxonomy for explainable artificial intelligence: a systematic survey of surveys on methods and concepts (2021, background): “– layers: The works of [144] and IIN [43] (invertible interpretation Concept completeness [144] IIN [43] networks) extend on the previous approaches and analyse a complete layer output space at once.”
Shape or Texture: Understanding Discriminative Features in CNNs (2021, methodology): “2.1 analyzed the consistency between the dimensionality estimate of (Esser et al., 2020) and that of (Bau et al., 2017).”
Network-to-Network Translation with Conditional Invertible Neural Networks (2020, methodology): “Recent works [76, 22] also utilize INNs to transform the latent distribution of an autoencoder to the base distribution.”
Making Sense of CNNs: Interpreting Deep Representations & Their Invariances with INNs (2020, methodology): “Instead of selecting a set of neurons directly, [15] learns an INN that transforms the original representation space to an interpretable space, where a single semantic concept is represented by a known group of neurons and a single neuron is involved in the encoding of just a single semantic concept”
iPOKE: Poking a Still Image for Controlled Stochastic Video Synthesis (2021, background): “Invertible neural networks (INNs) are learnable bijective functions often used to transform between two probability distributions, thus being a natural choice for addressing inverse problems [3], introspecting and explaining neural network representations [22, 32] and domain transfer [62, 63].”
Stochastic Image-to-Video Synthesis using cINNs (2021, background): “[22] Patrick Esser, Robin Rombach, and Björn Ommer.”
Explaining in Style: Training a GAN to explain a classifier in StyleSpace (2021, methodology): “Other explanation methods generate counterfactuals using attributes, where full or partial supervision for the desired attributes is available [10, 5].”