Understanding Representations Learned in Deep Architectures

By Dumitru Erhan, Aaron C. Courville, Yoshua Bengio. 2010

Subfields: Representation & Self-Supervised Learning, 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 Representation & Self-Supervised Learning

In Safety, Robustness & Interpretability

Citations in the corpus

Explaining nonlinear classification decisions with deep Taylor decomposition (2015, background): “transformations of the data [31], the role of particular neurons [32], or its global decision structure [33, 34].”
Evaluating the Visualization of What a Deep Neural Network Has Learned (2015, methodology): “Several methods have been developed to understand what a DNN has learned [17], [18].”
Practical Recommendations for Gradient-Based Training of Deep Architectures (2012, methodology): “The learning trajectories visualizations of Erhan et al. (2010b) have shown that even when starting from nearby conﬁgurations in function space, diﬀerent initializations seem to always fall in a diﬀerent eﬀective local minimum.”
Deep Learning for Health Informatics (2017, background): “1) Despite some recent work on visualizing high level features by using the weight filters in a CNN [141], [142], the entire deep learning model is often not interpretable.”
Explaining Predictions of Non-Linear Classifiers in NLP (2016, background): “…parallel, initiated by the computer vision domain, there is also a trend toward understanding deep learning models through visualization techniques (Erhan et al., 2010; Landecker et al., 2013; Zeiler and Fergus, 2014; Simonyan et al., 2014; Bach et al., 2015; Lapuschkin et al., 2016a) or through…”
Representation Learning: A Review and New Perspectives (2012, methodology): “…that layerwise stacking of feature extraction often yielded better representations, e.g., in terms of classification error (Larochelle et al., 2009; Erhan et al., 2010b), quality of the samples generated by a probabilistic model (Salakhutdinov and Hinton, 2009) or in terms of the invariance…”
Trojaning Attack on Neural Networks (2018, background): “It is highly difficult, if not impossible, to reason about or explain decisions made by a NN [26, 51].”
Auditing Deep Learning processes through Kernel-based Explanatory Models (2019, background): “…to improve the readability of a neural model in image-related tasks are based on back-ward algorithms that reuse arc weights to propagate the prediction down to the input (Erhan et al., 2010; Zeiler and Fergus, 2013), thus leading to the re-creation of meaningful patterns in the input space.”
Decoding the Deep: Exploring Class Hierarchies of Deep Representations Using Multiresolution Matrix Factorization (2017, background): “The seminal papers on deep networks [10, 9, 34] argue that, intuitively and empirically, higher layer representations learn abstract relationships between objects in a given image.”
Opportunities and Challenges in Explainable Artificial Intelligence (XAI): A Survey (2020, methodology): “[74], with focus on input patterns which maximize a given hidden unit activation.”
Learning to Deblur (2014, unknown): “is possible to calculate the gradient of the kernel prediction error of our NNs with respect to the input, we can optimize for the ideal input (similar to “activation maximization” [24]).”
Image denoising with multi-layer perceptrons, part 2: training trade-offs and analysis of their mechanisms (2012, methodology): “We answer this question using activation maximization , proposed by Erhan et al. (2010b).”
Visualizations of Deep Neural Networks in Computer Vision: A Survey (2017, methodology): “[16] 2010 Model quality assessment Pixel displays Representation Learning DBN MNIST, Caltech-101”
Applications of Deep Learning and Reinforcement Learning to Biological Data (2017, background): “This turns into a situation when the models are considered as ‘Black box’ [206].”
VisualBackProp: visualizing CNNs for autonomous driving (2016, methodology): “This technique cannot be applied to deep architectures based on auto-encoders as opposed to the subsequent work (Erhan et al., 2010; 2009).”
Mask-Specific Inpainting with Deep Neural Networks (2014, background): “Activation maximization [10] finds, for a given neuron in the last hidden layer, the respective input that maximizes that neuron’s activation, while constraining the input’s L2-norm.”
Explainable Artificial Intelligence: a Systematic Review (2020, methodology): “[222] and [223, 224] proposed two similar methods, based on Activation Maximization, which modify the input images in such a way to maximise the activation”