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

Abstract

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

arXiv comment29 pages, 35 figures, NIPS camera-ready

References in corpus (6)

• Auto-Encoding Variational Bayes
• Caffe: Convolutional Architecture for Fast Feature Embedding
• Understanding Neural Networks Through Deep Visualization
• Multifaceted Feature Visualization: Uncovering the Different Types of Features Learned By Each Neuron in Deep Neural Networks
• Generating Images with Perceptual Similarity Metrics based on Deep Networks
• Understanding Intra-Class Knowledge Inside CNN

Cited by in corpus (60)

• Deep learning with convolutional neural networks for EEG decoding and visualization
• Methods for Interpreting and Understanding Deep Neural Networks
• Conditional Image Synthesis With Auxiliary Classifier GANs
• Towards Explainable Artificial Intelligence
• A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications
• Places: An Image Database for Deep Scene Understanding
• An Introduction to Image Synthesis with Generative Adversarial Nets
• 3D Sketching using Multi-View Deep Volumetric Prediction
• From Clustering to Cluster Explanations via Neural Networks
• Generating and designing DNA with deep generative models
• CLIPDraw: Exploring Text-to-Drawing Synthesis through Language-Image Encoders
• Revisiting the Importance of Individual Units in CNNs via Ablation
• Photographic Image Synthesis with Cascaded Refinement Networks
• Overlearning Reveals Sensitive Attributes
• Deep learning as a tool for neural data analysis: speech classification and cross-frequency coupling in human sensorimotor cortex
• CVAE-GAN: Fine-Grained Image Generation through Asymmetric Training
• Testing and verification of neural-network-based safety-critical control software: A systematic literature review
• How convolutional neural network see the world - A survey of convolutional neural network visualization methods
• Be Your Own Prada: Fashion Synthesis with Structural Coherence
• Activation Maximization Generative Adversarial Nets
• Understanding trained CNNs by indexing neuron selectivity
• FaceLift: A transparent deep learning framework to beautify urban scenes
• Visually-Aware Fashion Recommendation and Design with Generative Image Models
• Design by adaptive sampling
• Scribbler: Controlling Deep Image Synthesis with Sketch and Color
• Latent Constraints: Learning to Generate Conditionally from Unconditional Generative Models
• Adversarial Message Passing For Graphical Models
• TIP: Typifying the Interpretability of Procedures
• Zero-Shot Text-Guided Object Generation with Dream Fields
• Quantitative Evaluations on Saliency Methods: An Experimental Study
• Synthetic to Real Adaptation with Generative Correlation Alignment Networks
• SUNLayer: Stable denoising with generative networks
• Generalizing multistain immunohistochemistry tissue segmentation using one-shot color deconvolution deep neural networks
• DeepMasterPrints: Generating MasterPrints for Dictionary Attacks via Latent Variable Evolution
• A Game-Theoretic Taxonomy of Visual Concepts in DNNs
• Learning from Higher-Layer Feature Visualizations
• Influence-Directed Explanations for Deep Convolutional Networks
• Decoupled Learning for Conditional Adversarial Networks
• Why do These Match? Explaining the Behavior of Image Similarity Models
• "I know it when I see it". Visualization and Intuitive Interpretability
• Visualizing Transfer Learning
• Deeper, Broader and Artier Domain Generalization
• Interpreting Face Inference Models using Hierarchical Network Dissection
• Depth Structure Preserving Scene Image Generation
• Diverse feature visualizations reveal invariances in early layers of deep neural networks
• Updating the generator in PPGN-h with gradients flowing through the encoder
• Measure Utility, Gain Trust: Practical Advice for XAI Researcher
• Building a Telescope to Look Into High-Dimensional Image Spaces
• Understanding Regularization to Visualize Convolutional Neural Networks
• Evolution of Images with Diversity and Constraints Using a Generator Network
• Text-to-image Synthesis via Symmetrical Distillation Networks
• A Chain Graph Interpretation of Real-World Neural Networks
• When and where do feed-forward neural networks learn localist representations?
• Diagnostic Visualization for Deep Neural Networks Using Stochastic Gradient Langevin Dynamics
• Image Score: How to Select Useful Samples
• Efficient Modelling Across Time of Human Actions and Interactions
• Model Reduction of Shallow CNN Model for Reliable Deployment of Information Extraction from Medical Reports
• Semantic Network Interpretation
• Collaborative creativity with Monte-Carlo Tree Search and Convolutional Neural Networks
• Train, Diagnose and Fix: Interpretable Approach for Fine-grained Action Recognition