Research map / Label-efficient and robust learning / Recognition and long-tail
Human-Guided Complexity-Controlled Abstractions
Andi Peng, Mycal Tucker, Eoin M. Kenny, Noga Zaslavsky, Pulkit Agrawal, Julie Shah
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 |
|---|---|---|---|
| Use VQ-VIB_C instead of beta-VAE or VQ-VIB_N encoders | finetuning encoder · FashionMNIST 3-way, k=1 | accuracy: 40% and 55% -> 70% | Improved |
| Reduce representation complexity (higher MSE) up to a point | VQ-VIB_C · FashionMNIST 3-way, k=1 | accuracy: MSE < 0.055 positive correlation, worse beyond | Mixed |
| Use VQ-VIB_C instead of beta-VAE | finetuning encoder · CIFAR100 2-way, k=1 | accuracy: 50% -> 70% | Improved |
| Increase finetuning examples per class from 1 to 50 | encoders · CIFAR100 2-way | accuracy gap | No clear effect |
| Compress representations on 20-way task | VQ-VIB_C · CIFAR100 20-way, k=1 | accuracy | Hurt |
| Increase number of combined quantized vectors n (1 to 2, 4) (ablation) | VQ-VIB_C · FashionMNIST, iNat | finetuning accuracy: n=1 best | Hurt |
| Penalize entropy instead of complexity (ablation) | VQ-VIB_C and VQ-VIB_N · FashionMNIST | finetuning accuracy | Improved |
| Use VQ-VIB_C instead of VQ-VIB_N under entropy regularization (ablation) | encoder · FashionMNIST | finetuning accuracy | Improved |
| Choose most complex encoder when finetuning data is large (k=50) | encoder selection · FashionMNIST, CIFAR100, iNat | accuracy | No clear effect |
| Select encoder by validation-set accuracy with small k and v | encoder selection · FashionMNIST k=2, v=1 | accuracy: 73% vs over 80% | Hurt |
| Increase validation set and k | encoder selection · FashionMNIST | accuracy: 0.73 -> 0.97 | Improved |
| Increase finetuning data to 1010-way iNat with large k | VQ-VIB_C vs beta-VAE · iNat 1010-way | accuracy | Hurt |
Closest papers by meaning
- Disentangling Representations through Multi-task Learning (ICLR 2025 · similarity 0.86)
- Geometry of naturalistic object representations in recurrent neural network models of working memory (NeurIPS 2024 · similarity 0.86)
- Turning large language models into cognitive models (ICLR 2024 · similarity 0.85)
- Attention as a Hypernetwork (ICLR 2025 · similarity 0.85)
- Evaluating alignment between humans and neural network representations in image-based learning tasks (NeurIPS 2024 · similarity 0.85)
- Scaling can lead to compositional generalization (NeurIPS 2025 · 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.
