Research map / Architectures and efficiency / RNNs and normalization
STCN: Stochastic Temporal Convolutional Networks
Emre Aksan, Otmar Hilliges
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| connect all hierarchical latent samples directly to the output instead of using only the lowest-level sample (ablation) | STCN · IAM-OnDB | ELBO: ≥ 1339 -> ≥ 1796 | Improved |
| connect all hierarchical latent samples directly to the output instead of using only the lowest-level sample (ablation) | STCN · Deepwriting | average log-likelihood per sequence: ≥ 605 -> ≥ 797 | Improved |
| connect all hierarchical latent samples directly to the output instead of using only the lowest-level sample (ablation) | STCN · TIMIT | ELBO: ≥ 69195 -> ≥ 71386 | Improved |
| connect all hierarchical latent samples directly to the output instead of using only the lowest-level sample (ablation) | STCN · Blizzard | average log-likelihood per sequence: 15800 -> 16288 | Improved |
| replace TCN blocks with LSTM cells while retaining the latent hierarchy and dense connections (ablation) | STCN-dense · TIMIT | ELBO: ≥ 71386 -> ≥ 50113 | Hurt |
| replace TCN blocks with LSTM cells while retaining the latent hierarchy and dense connections (ablation) | STCN-dense · IAM-OnDB | ELBO: ≥ 1796 -> ≥ 1581 | Hurt |
| replace TCN blocks with LSTM cells in the 25x256 configuration (ablation) | STCN-dense · TIMIT | ELBO: ≥ 71386 -> ≥ 26240 | Hurt |
| replace TCN blocks with LSTM cells in the 25x256 configuration (ablation) | STCN-dense · IAM-OnDB | ELBO: ≥ 1796 -> ≥ 1377 | Hurt |
| replace concatenation of latent samples with summation before the output (ablation) | STCN-dense · sequence modeling experiments | ELBO | Hurt |
| increase speech-model convolution filters from 256 to 512 (ablation) | STCN-dense · TIMIT | average log-likelihood per sequence: ≥ 71386 -> ≥ 77438 | Improved |
| increase speech-model convolution filters from 256 to 512 (ablation) | STCN-dense · Blizzard | average log-likelihood per sequence: ≥ 16288 -> ≥ 17670 | Improved |
Closest papers by meaning
- Long-Term Recurrent Convolutional Networks for Visual Recognition and Description (TPAMI 2016 · similarity 0.89)
- Combining Recurrent, Convolutional, and Continuous-time Models with Linear State Space Layers (NeurIPS 2021 · similarity 0.87)
- Making Convolutional Networks Recurrent for Visual Sequence Learning (CVPR 2018 · similarity 0.87)
- Temporal Deformable Convolutional Encoder-Decoder Networks for Video Captioning (AAAI 2019 · similarity 0.87)
- StNet: Local and Global Spatial-Temporal Modeling for Action Recognition (AAAI 2019 · similarity 0.87)
- Universal Approximation of Input-Output Maps by Temporal Convolutional Nets (NeurIPS 2019 · similarity 0.86)
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.
