Visualizing Higher-Layer Features of a Deep Network

By Dumitru Erhan, Yoshua Bengio, Aaron C. Courville, Pascal Vincent. 2009

1,454 citations recorded by the source corpus.

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

Intriguing properties of neural networks (2013, background): “So far, unit-level inspection methods had relatively little utility beyond conﬁrming certain intuitions regarding the complexity of the representations learned by a deep neural network [6, 13, 7, 4].”
Visualizing and Understanding Convolutional Networks (2013, background): “(Erhan et al., 2009) ﬁnd the optimal stimulus for each unit by performing gradient descent in image space to maximize the unit’s activation.”
On the Anatomy of MCMC-based Maximum Likelihood Learning of Energy-Based Models (2019, background): “Previous AM research focuses on maximizing the response of pre-trained classifier networks [7, 24, 26, 29].”
CLIPDraw: Exploring Text-to-Drawing Synthesis through Language-Image Encoders (2021, background): “This framework is often referred to as activation maximization (Erhan et al., 2009; Nguyen et al., 2016; Mordvintsev et al., 2015), where a random image is optimized through backpropogation to increase certain neuron activations of a pretrained network.”
High Fidelity Visualization of What Your Self-Supervised Representation Knows About (2021, neutral): “Deterministic visualization methods: Many early works (Erhan et al., 2009; Zeiler & Fergus, 2014; Simonyan et al., 2013; Selvaraju et al., 2016; Smilkov et al., 2017) used gradient based techniques to visualize what is learned by neural networks.”
A critical analysis of self-supervision, or what we can learn from a single image (2019, methodology): “In order to understand what deeper neurons are responding to in our model, we visualize random neurons via activation maximization (Erhan et al., 2009; Zeiler & Fergus, 2014) in each layer.”
Towards the Science of Security and Privacy in Machine Learning (2016, background): “An approach named activation maximization synthesizes an input that highly activates a specific neuron in a neural network [113].”
Feature Purification: How Adversarial Training Performs Robust Deep Learning (2020, background): “Visualizing the features presented by deeper convolutional layers is an active research area, dating back at least to [28].”
Transformer Feed-Forward Layers Are Key-Value Memories (2020, methodology): “Characterizing the functionality of memory cells based on examples that trigger maximal activations has been used previously in NLP (Rethmeier et al., 2020) and vision (Erhan et al., 2009).”
Axiomatic Attribution for Deep Networks (2017, background): “Most of this work has 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; Yosi”
Building high-level features using large scale unsupervised learning (2011, methodology): “The second approach is to perform numerical optimization to ﬁnd the optimal stimulus (Berkes & Wiskott, 2005; Erhan et al., 2009; Le et al., 2010).”
Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps (2013, methodology): “First, we demonstrate that understandable visualisations of ConvNet classiﬁcation models can be obtained using the numerical optimisation of the input image [5] (Sect.”
Deep Clustering for Unsupervised Learning of Visual Features (2018, background): “We assess the quality of a target filter by learning an input image that maximizes its activation [18, 70].”
Deep neural networks are easily fooled: High confidence predictions for unrecognizable images (2014, methodology): “(2) It is possible to find highconfidence images (pentagon) using our directly encoded EA or gradient ascent optimization starting from a random or blank image (I0) [11, 13, 26].”
Understanding deep image representations by inverting them (2014, methodology): “The latter was used in [6, 26] to generate images representative of given neurons.”
Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization (2016, methodology): “Other visualization methods synthesize images to maximally activate a network unit [51,16] or invert a latent representation [41,15].”
Synthesizing the preferred inputs for neurons in neural networks via deep generator networks (2016, background): “Synthesizing preferred stimuli is called activation maximization [4–8, 3, 9].”
Understanding Neural Networks Through Deep Visualization (2015, background): “For example, Erhan et al. (2009) synthesized images that cause high activations for particular units.”