Inverting Convolutional Networks with Convolutional Networks

By Alexey Dosovitskiy, Thomas Brox. 2015

114 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition

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

Citations in the corpus

What makes ImageNet good for transfer learning? (2016, methodology): “Another line of prior work investigated questions such as how distributed are CNN representations [2, 43], what is the the effect of feature transformations such as binarization [2, 14], how do CNN features change with systematic variations in scene factors [3], how invariant are features to differe”
Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization (2016, methodology): “Other visualization methods synthesize images to maximally activate a network unit [40, 11] or invert a latent representation [31, 10].”
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.”
DISCO: Dynamic and Invariant Sensitive Channel Obfuscation for deep neural networks (2020, positive): “This attack scheme is inspired from [33, 53, 54].”
A Unified Gradient Regularization Family for Adversarial Examples (2015, background): “While such assumption might not hold strictly in more complex data or data in other domains, there are researches indicate that the encoding of such information exist [13].”
A Survey on Explainable Artificial Intelligence (XAI): Toward Medical XAI (2019, methodology): “t statistically [65] 7 X + Attention map with autofocus convolutional layer [66] 7 X DeconvNet [71] 7 X Inversion Signal Inverting representation with natural image prior [72] 7 X Inversion using CNN [73] 7 X Guided backpropagation [74], [90] 7 X Activation maximization/optimization [37] 7 X Optimiz”
Evaluating the Visualization of What a Deep Neural Network Has Learned (2015, background): “een developed to understand what a DNN has learned [16], [17], [18]. While in DNN a large body of work is dedicated to visualize particular neurons or neuron layers [1], [19], [20], [21], [22], [23], [24], we focus here on methods which visualize the impact of particular regions of a given and ﬁxed ”
Deep Image Prior (2017, methodology): “The reconstructions obtained with the deep image prior are in many ways at least as natural as [8], yet they are not biased by the learning process.”
Visualizing Deep Convolutional Neural Networks Using Natural Pre-images (2015, background): “A related line of work (Dosovitskiy and Brox 2015; Bishop 1995) is to learn a second neural network to act as the inverse of the original one.”
Understanding and Improving Convolutional Neural Networks via Concatenated Rectified Linear Units (2016, methodology): “Unlike other reconstruction methods (Dosovitskiy & Brox, 2015; Mahendran & Vedaldi, 2015), our simple linear reconstruction algorithm does not involve any additional learning.”
Fast Patch-based Style Transfer of Arbitrary Style (2016, background): “We first present this operation at a high level, followed by more details on our implementation.”
Unsupervised Learning on Neural Network Outputs: With Application in Zero-Shot Learning (2015, background): “onal costs. 6. Discussion and Conclusion The outputs of a neural network contains rich information. It has been shown that one can reconstruct the whole image to some degree with only its CNN outputs [9]. And smooth regularization on the output distribution of a neural network can help in reducing g”
Learning Aligned Cross-Modal Representations from Weakly Aligned Data (2016, unknown): “We refer readers to [7] for full details.”
Visualizing and Understanding Deep Texture Representations (2015, background): “Dosovitskiy and Brox [8] propose a “deconvolutional network” to invert a CNN in a feed-forward manner but the approach tends to produce blurry images since the inverse is not uniquely defined.”
Deep Exemplar 2D-3D Detection by Adapting from Real to Rendered Views (2015, background): “Looking at our top false positives in Figure 3 confirms this, since we find many parts of empty bookcases or bookcases from other categories.”
Layer-Wise Relevance Propagation for Deep Neural Network Architectures (2016, background): “Other approaches deal with inverting neural network representations [8, 16] or are about finding images with unusual properties and wrong classification outcomes [19, 11, 24].”
Unsupervised Learning of Visual Structure using Predictive Generative Networks (2015, methodology): “We use a model consisting of a convolutional neural network (CNN), an LSTM, and a deconvolutional neural network (deCNN) (Dosovitskiy & Brox, 2015; Kulkarni et al., 2015).”
Relief Impression Image Detection : Unsupervised Extracting Objects Directly From Feature Arrangements of Deep CNN (2016, result): “This result is coincide to the CNN visualization research [20], [23], [24], those works find that the higher convolutional layer keeps less location informantion.”