Synthesizing the preferred inputs for neurons in neural networks via\n deep generator networks

2016/05/30 by Anh Son Nguyen, Nguyen, Anh, Alexey Dosovitskiy +8 · 39 citations
Biochemistry, Genetics and Molecular Biology · Computer Science · #Cell Image Analysis Techniques #Neural Networks and Applications #Advanced Neural Network Applications

paper · pdf · doi:10.48550/arxiv.1605.09304

Abstract

Deep neural networks (DNNs) have demonstrated state-of-the-art results on\nmany pattern recognition tasks, especially vision classification problems.\nUnderstanding the inner workings of such computational brains is both\nfascinating basic science that is interesting in its own right - similar to why\nwe study the human brain - and will enable researchers to further improve DNNs.\nOne path to understanding how a neural network functions internally is to study\nwhat each of its neurons has learned to detect. One such method is called\nactivation maximization (AM), which synthesizes an input (e.g. an image) that\nhighly activates a neuron. Here we dramatically improve the qualitative state\nof the art of activation maximization by harnessing a powerful, learned prior:\na deep generator network (DGN). The algorithm (1) generates qualitatively\nstate-of-the-art synthetic images that look almost real, (2) reveals the\nfeatures learned by each neuron in an interpretable way, (3) generalizes well\nto new datasets and somewhat well to different network architectures without\nrequiring the prior to be relearned, and (4) can be considered as a\nhigh-quality generative method (in this case, by generating novel, creative,\ninteresting, recognizable images).\n

Citations

• Auto-Encoding Variational Bayes
• Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

Cited by

• Localising Shortcut Learning in Pixel Space via Ordinal Scoring Correlations for Attribution Representations (OSCAR)
• VFMF: World Modeling by Forecasting Vision Foundation Model Features
• Back to the Baseline: Examining Baseline Effects on Explainability Metrics
• Zero-Shot Textual Explanations via Translating Decision-Critical Features
• TIE: A Training-Inversion-Exclusion Framework for Visually Interpretable and Uncertainty-Guided Out-of-Distribution Detection
• TRANSPORTER: Transferring Visual Semantics from VLM Manifolds
• Towards Open-Ended Visual Scientific Discovery with Sparse Autoencoders
• Scribbler: Controlling Deep Image Synthesis with Sketch and Color
• DenoGrad: A Gradient-Based Framework for Data Refinement in Tabular and Time-Series Learning
• Learning from Higher-Layer Feature Visualizations
• An Introduction to Image Synthesis with Generative Adversarial Nets
• Places: An Image Database for Deep Scene Understanding
• Methods for Interpreting and Understanding Deep Neural Networks
• Revisiting the Importance of Individual Units in CNNs via Ablation
• DEXTER: Diffusion-Guided EXplanations with TExtual Reasoning for Vision Models
• Semantic Network Interpretation
• EXP-CAM: Explanation Generation and Circuit Discovery Using Classifier Activation Matching
• Activation Matching for Explanation Generation
• Generalizing multistain immunohistochemistry tissue segmentation using one-shot color deconvolution deep neural networks
• Quantitative Evaluations on Saliency Methods: An Experimental Study
• Semantic Understanding of Scenes Through the ADE20K Dataset
• Generating and designing DNA with deep generative models
• Train, Diagnose and Fix: Interpretable Approach for Fine-grained Action Recognition
• Depth Structure Preserving Scene Image Generation
• "I know it when I see it". Visualization and Intuitive Interpretability
• Influence-Directed Explanations for Deep Convolutional Networks
• A Chain Graph Interpretation of Real-World Neural Networks
• Zero-Shot Text-Guided Object Generation with Dream Fields
• Design by adaptive sampling
• Model Interpretability and Rationale Extraction by Input Mask Optimization
• Evolution of Images with Diversity and Constraints Using a Generator Network
• SUNLayer: Stable denoising with generative networks
• Collaborative creativity with Monte-Carlo Tree Search and Convolutional Neural Networks
• Image Score: How to Select Useful Samples
• GENIAL: Generative Design Space Exploration via Network Inversion for Low Power Algorithmic Logic Units
• Understanding Regularization to Visualize Convolutional Neural Networks
• Model Reduction of Shallow CNN Model for Reliable Deployment of Information Extraction from Medical Reports
• A Game-Theoretic Taxonomy of Visual Concepts in DNNs
• When and where do feed-forward neural networks learn localist representations?

Related

• Understanding Neural Networks via Feature Visualization: A survey
• Representation Understanding via Activation Maximization
• Multifaceted Feature Visualization: Uncovering the Different Types of Features Learned By Each Neuron in Deep Neural Networks
• DISCOVER: Making Vision Networks Interpretable via Competition and Dissection
• NeuroGen: activation optimized image synthesis for discovery neuroscience