Research map / Label-efficient and robust learning / Recognition and long-tail
Variational Autoencoder for Deep Learning of Images, Labels and Captions
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 |
|---|---|---|---|
| stochastic unpooling instead of deterministic unpooling (ablation) | DGDN-CNN VAE · MNIST, CIFAR-10, CIFAR-100, Caltech 101, Caltech 256 | classification test error (%): e.g. CIFAR-10 10.74 -> 8.19; Caltech 101 14.79 -> 11.99 | Helped |
| CNN-based recognition model (encoder) replacing Gibbs/MCEM inference | DGDN · Benchmark classification (Table 1) | test time per image (ms): up to 400x faster at test time; e.g. Caltech 256 8.9 ms (MCEM) -> 0.3 ms | Helped |
| CNN-based recognition model replacing Gibbs/MCEM inference | DGDN · Benchmark classification (Table 1) | classification error (%): comparable accuracy; better on several datasets, slightly worse on MNIST (0.38 vs 0.37 Gibbs) and CIFAR-100 (35.01 vs 34 | No clear effect |
| joint image-caption training versus two-step (unsupervised encoder then RNN) (ablation) | DGDN-CNN VAE + GRU captioning · Flickr8k, Flickr30k, MS COCO | BLEU-4 / CIDEr: MS COCO CIDEr 0.58 -> 0.89; Flickr8k BLEU-4 0.17 -> 0.22 | Helped |
| semi-supervised use of ImageNet 2012 training images as uncaptioned data in joint captioning (ablation) | Joint captioning model · Flickr8k, Flickr30k, MS COCO | BLEU-4 / CIDEr: Flickr8k BLEU-4 0.22 -> 0.25; MS COCO CIDEr 0.89 -> 0.90 (gain largest on small datasets) | Helped |
| semi-supervised training with 10 labels/class versus unsupervised (TSVM on features) (ablation) | DGDN-CNN VAE on MNIST · MNIST, 10 labels per class | test error (%): 5.83 -> 1.49 | Helped |
| Bayesian SVM replaced with softmax classifier (ablation) | Semi-supervised ImageNet classification · ImageNet 2012 semi-supervised | accuracy: softmax slightly worse by ~1% (labels <30%) and ~0.5% (labels >30%) | Hurt |
| deterministic unpooling replacing stochastic unpooling in ImageNet 2012 setting (ablation) | DGDN-CNN VAE · ImageNet 2012 | top-1 accuracy: not reported; described as much worse | Hurt |
Closest papers by meaning
- Improving Image Captioning with Conditional Generative Adversarial Nets (AAAI 2019 · similarity 0.85)
- Regularizing RNNs for Caption Generation by Reconstructing the Past With the Present (CVPR 2018 · similarity 0.85)
- Improved Variational Autoencoders for Text Modeling using Dilated Convolutions (ICML 2017 · similarity 0.84)
- Gradient Origin Networks (ICLR 2021 · similarity 0.84)
- Regularizing Deep Networks by Modeling and Predicting Label Structure (CVPR 2018 · similarity 0.84)
- Variational Autoencoder for Semi-Supervised Text Classification (AAAI 2017 · similarity 0.84)
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.