Research map / Video / Video segmentation and tracking
Deep Predictive Coding Networks for Video Prediction and Unsupervised Learning
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 predictive coding with errors passed between layers rather than the CNN-LSTM encoder-decoder activation pathway | PredNet · Rotating Faces test set | MSE: 0.0180 -> 0.0152 | Helped |
| apply prediction loss to upper layers with weight 0.1 instead of concentrating loss at the lowest layer (ablation) | PredNet · Rotating Faces test set | MSE: 0.0152 -> 0.0157 | Hurt |
| use predictive coding with errors passed between layers rather than the CNN-LSTM encoder-decoder activation pathway | PredNet · CalTech Pedestrian test set | MSE: 3.67e-3 -> 3.13e-3 | Helped |
| apply prediction loss to upper layers with weight 0.1 instead of concentrating loss at the lowest layer (ablation) | PredNet · CalTech Pedestrian test set | MSE: 3.13e-3 -> 3.33e-3 | Hurt |
| use split positive and negative prediction-error populations instead of linear error responses (ablation) | PredNet · CalTech Pedestrian test set | MSE (x 10^-3): 3.20 -> 3.13 | Helped |
| double the activation filters to match PredNet's filter count (ablation) | CNN-LSTM Enc.-Dec. · CalTech Pedestrian test set | MSE (x 10^-3): 3.67 -> 3.82 | Hurt |
| pass the lowest-layer error E0 while retaining activation passing at higher layers (ablation) | CNN-LSTM Enc.-Dec. · CalTech Pedestrian test set | MSE (x 10^-3): 3.67 -> 3.41 | Helped |
| split activations into positive and negative populations before passing them onward (ablation) | CNN-LSTM Enc.-Dec. · CalTech Pedestrian test set | MSE (x 10^-3): 3.67 -> 3.71 | Hurt |
| use PredNet's error-passing architecture rather than DFN | PredNet · 64x64 Car-Cam Dataset | MSE (per-pixel): 1.71e-3 -> 1.16e-3 | Helped |
| use PredNet rather than DNA for next-frame prediction | PredNet · Human3.6M | PSNR: 42.1 -> 38.9 | Hurt |
| train with loss on upper layers at weight 0.1 rather than only the lowest layer (ablation) | PredNet · Comma.ai steering angle estimation | MSE | Hurt |
Closest papers by meaning
- Anticipating Visual Representations From Unlabeled Video (CVPR 2016 · similarity 0.90)
- Deep Predictive Coding Network with Local Recurrent Processing for Object Recognition (NeurIPS 2018 · similarity 0.89)
- Predify: Augmenting deep neural networks with brain-inspired predictive coding dynamics (NeurIPS 2021 · similarity 0.89)
- Decomposing Motion and Content for Natural Video Sequence Prediction (ICLR 2017 · similarity 0.89)
- Deep Predictive Coding Network for Object Recognition (ICML 2018 · similarity 0.88)
- Learning from Unlabelled Videos Using Contrastive Predictive Neural 3D Mapping (ICLR 2020 · similarity 0.88)
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.