Research map / Learning theory and algorithms / Prediction and decision losses
Object Representations as Fixed Points: Training Iterative Refinement Algorithms with Implicit Differentiation
Michael Chang, Thomas L. Griffiths, Sergey Levine
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 |
|---|---|---|---|
| Train slot attention with implicit differentiation rather than backpropagating through unrolled iterations | SLATE on CLEVR-Mirror 64x64 · CLEVR-Mirror 64x64 | cross entropy: around 130 -> 10x lower cross entropy loss | Improved |
| Use implicit differentiation for slot attention | SLATE · CLEVR test | FID: 25.89 -> 22.19 | Improved |
| Use implicit differentiation for slot attention | SLATE · CLEVR test | MSE: 67.04 -> 10.66 | Improved |
| Use implicit differentiation for slot attention | SLATE · COCO test | FID: 147.48 -> 127.79 | Improved |
| Use implicit differentiation for slot attention | SLATE · COCO test | MSE: 1821.75 -> 1659.15 | Improved |
| Use implicit differentiation for slot attention | SLATE · ShapeStacks test | FID: 34.76 -> 34.2 | Improved |
| Use implicit differentiation for slot attention | SLATE · ShapeStacks test | MSE: 312.14 -> 108.67 | Improved |
| Use the IB configuration (iteration forward solver with Broyden gradient estimate) (ablation) | Slot attention · CLEVR-Mirror 64x64 | optimization stability | Hurt |
| Replace the seven-iteration vanilla model with one iteration (ablation) | Vanilla slot attention · CLEVR | reconstruction MSE: 2x worse reconstruction MSE than implicit slot attention with seven iterations | Hurt |
| Remove gradient clipping (ablation) | Vanilla SLATE · Training | training stability | Hurt |
| Remove gradient clipping (ablation) | Implicit SLATE · Training | training stability | No clear effect |
| Remove learning-rate warmup (ablation) | Vanilla SLATE · Training | training stability | Hurt |
Closest papers by meaning
- Object centric architectures enable efficient causal representation learning (ICLR 2024 · similarity 0.86)
- Object-Centric Pretraining via Target Encoder Bootstrapping (ICLR 2025 · similarity 0.86)
- Learning to Compose: Improving Object Centric Learning by Injecting Compositionality (ICLR 2024 · similarity 0.85)
- Unlocking Compositional Generalization in Continual Few-Shot Learning (NeurIPS 2026 · similarity 0.85)
- Slot-Guided Adaptation of Pre-trained Diffusion Models for Object-Centric Learning and Compositional Generation (ICLR 2025 · similarity 0.85)
- Cycle Consistency Driven Object Discovery (ICLR 2024 · 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.
