Understanding deep image representations by inverting them

By Aravindh Mahendran, Andrea Vedaldi. 2014

2,100 citations recorded by the source corpus.

Image representations, from SIFT and Bag of Visual Words to Convolutional Neural Networks (CNNs), are a crucial component of almost any image understanding system. Nevertheless, our understanding of them remains limited. In this paper we conduct a direct analysis of the visual information contained in representations by asking the following question: given an encoding of an image, to which extent is it possible to reconstruct the image itself? To answer this question we contribute a general framework to invert representations. We show that this method can invert representations such as HOG more accurately than recent alternatives while being applicable to CNNs too. We then use this technique to study the inverse of recent state-of-the-art CNN image representations for the first time. Among our findings, we show that several layers in CNNs retain photographically accurate information about the image, with different degrees of geometric and photometric invariance.

Subfields: Computer Vision and Pattern Recognition, Safety, Robustness & Interpretability

How researchers see this paper

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

In Computer Vision and Pattern Recognition

In Safety, Robustness & Interpretability

Citations in the corpus

Adversarial Examples Are Not Bugs, They Are Features (2019, background): “It also suggests that approaches aiming to enhance the interpretability of a given model by enforcing “priors” for its explanation [MV15; OMS17; Smi+17] actually hide features that are “meaningful” and predictive to standard models.”
Making Convolutional Networks Shift-Invariant Again (2019, background): “…hidden units (Girshick et al., 2014; Zhou et al., 2015), actively maximizing hidden units (Mordvintsev et al., 2015), and mapping features back into pixel space (Zeiler & Fergus, 2014; H´enaff & Si-moncelli, 2016; Mahendran & Vedaldi, 2015; Dosovitskiy & Brox, 2016a;b; Nguyen et al., 2017).”
Are GANs Created Equal? A Large-Scale Study (2017, methodology): “we solve z = argminz∈Rdz ||x̂−G(z)||(2)2, using gradient descent on z while keeping G fixed [17].”
Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles (2016, background): “Some recent work has devoted efforts towards the visualization of CNNs to better understand how they work and how we can exploit them [37,30,24,19].”
Generating Images with Perceptual Similarity Metrics based on Deep Networks (2016, methodology): “Note that similar to Dosovitskiy & Brox (2015) and unlike Mahendran & Vedaldi (2015), our method does not require the feature representation being inverted to be differentiable.”
Semantic Image Inpainting with Deep Generative Models (2016, methodology): “Additionally, back-propagation has also been used to visualize and understand the learned features in a trained network, by “inverting” the network through updating the gradient at the input layer [26, 5, 35, 21].”
Deep multi-scale video prediction beyond mean square error (2015, background): “Our GDL is fundamentally different: In (Mahendran & Vedaldi, 2015), the total variation takes only the reconstructed frame in input, whereas our loss penalises gradient differences between the prediction and the true output.”
Feature Purification: How Adversarial Training Performs Robust Deep Learning (2020, background): “ave tried to various ways to resolve this issue (see e.g. an extensive survey by Olah et al. [76] and the references therein). At a high level, some penalizes the image to remove high-frequency noise [66, 70, 72, 79, 98]; some searches for images that can still 11We used the implementations from htt”
Axiomatic Attribution for Deep Networks (2017, background): “…been on networks trained on computer vision tasks, and deals with understanding what a speciﬁc neuron computes (Erhan et al., 2009; Le, 2013) and interpreting the representations captured by neurons during a prediction (Mahendran & Vedaldi, 2015; Dosovitskiy & Brox, 2015; Yosinski et al., 2015).”
Perceptual Losses for Real-Time Style Transfer and Super-Resolution (2016, methodology): “As demonstrated in [7] and reproduced in Fig.”
Unsupervised Cross-Domain Image Generation (2016, background): “The last loss, L TV is an anisotropic total variation loss (Rudin et al., 1992; Mahendran & Vedaldi, 2015), which is added in order to slightly smooth the resulting image.”
Precomputed Real-Time Texture Synthesis with Markovian Generative Adversarial Networks (2016, background): “1 is a smoothness prior for pixels [23].”
Learning Deep Features for Discriminative Localization (2015, background): “By removing the fully-connected layers and retaining most of the performance, we are able to understand our network from the beginning to the end.”
Inverting Visual Representations with Convolutional Networks (2015, methodology): “For lower layers, the reconstruction error of [19] is still much higher than of our method, even though visually the images look somewhat sharper.”
Understanding Neural Networks Through Deep Visualization (2015, background): “One approach is to study each layer as a group and investigate the type of computation performed by the set of neurons on a layer as a whole (Yosinski et al., 2014; Mahendran & Vedaldi, 2014).”
Visualizing and Understanding Neural Models in NLP (2015, background): “(1) Inversion: Inverting the representations by training an additional model to project outputs from different neural levels back to the initial input images (Mahendran and Vedaldi, 2014; Vondrick et al., 2013; Weinzaepfel et al., 2011).”
Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI (2019, neutral): “One of the seminal works in the second category is [137] .”
DISCO: Dynamic and Invariant Sensitive Channel Obfuscation for deep neural networks (2020, positive): “This attack scheme is inspired from [33, 53, 54].”