Research map / Image generation / Autoregressive and latent generation
Synthesizing the preferred inputs for neurons in neural networks via deep generator networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a learned deep generator network prior and optimize its input code instead of directly optimizing image pixels | Activation maximization for CaffeNet neurons · ImageNet | visualization quality | Helped |
| Use the fc6 code prior rather than conv3, conv5, or fc7 code priors (ablation) | DGN-AM · CaffeNet neurons | coherent global structure | Helped |
| Optimize convolutional-layer codes rather than fully connected-layer codes (ablation) | DGN-AM · CaffeNet neurons | coherent global structure | Hurt |
| Clip optimized code units to the activation range [0, 3σ] (ablation) | DGN-AM · Neuron visualizations | image quality | Helped |
| Use an ImageNet-trained prior to visualize a different dataset with a matching CaffeNet architecture | DGN-AM · MIT Places | visualization quality | Helped |
| Use the CaffeNet-trained prior for a heavily modified target architecture | DGN-AM · UCF-101 activity classes | visualization quality | Hurt |
| Apply the CaffeNet-trained prior to target architectures increasingly different from CaffeNet | DGN-AM · ImageNet neurons in GoogLeNet and ResNet | visualization quality | Hurt |
| Use the DGN prior rather than total variation and mean-image initialization | Hidden-neuron activation maximization · CaffeNet | realism of color and texture | Helped |
| Use the same ImageNet-trained prior on neurons trained with cut-up, BRG-color, or blurred images | DGN-AM · Modified ImageNet classes | reflection of learned image properties | Helped |
| Train the target network on mostly canonical examples of a class | DGN-AM visualizations · ImageNet canonical-versus-noncanonical class pairs | canonical appearance | Helped |
| Train the target network to overfit using only 10% or 25% of ImageNet images | DGN-AM · ImageNet output neurons | visualization quality: 20.2% and 31.5% validation accuracy versus 57.4% | Hurt |
| Shuffle frames rather than preserve their order | LRCN activity classifier · UCF-101 | classification decision | No clear effect |
Closest papers by meaning
- Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space (CVPR 2017 · similarity 0.89)
- From Activation to Initialization: Scaling Insights for Optimizing Neural Fields (CVPR 2024 · similarity 0.87)
- Exemplary Natural Images Explain CNN Activations Better than State-of-the-Art Feature Visualization (ICLR 2021 · similarity 0.87)
- Adversarial Attacks on the Interpretation of Neuron Activation Maximization (AAAI 2024 · similarity 0.86)
- Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data (ICLR 2024 · similarity 0.85)
- Dynamic Neural Regeneration: Enhancing Deep Learning Generalization on Small Datasets (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.