Research map / Probabilistic and scientific modeling / Causal discovery
Object centric architectures enable efficient causal representation learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that injective-generator assumptions in causal representation learning break for multi-object scenes, and proposes an object-centric (SA-MESH slot attention) encoder with sparse-perturbation weak supervision and slot matching that disentangles per-object properties with fewer perturbations.
Problem
Causal representation learning assumes an injective generative function over d-dimensional latent vectors, but when observations contain unordered sets of indistinguishable objects this fails, so disentanglement of object properties breaks down.
Approach
Treat multi-object scenes as sets: use an object-centric slot-attention encoder (SA-MESH) that partitions the image into per-object patches with a shared per-slot encoder, then apply sparse-perturbation weak supervision with a matching step to find the perturbed slot. This reduces the multi-object problem to single-object disentanglement and needs about k times fewer perturbations.
Key results
- On 2D shapes with unknown fully sparse perturbations, the proposed method reaches MCC 1.00 ± 0.01 for (pos_x, pos_y) with n=2 and 3 objects, and 0.95 ± 0.01 for five properties with n=2.
- The non-injective ResNet18 (CNN) baseline drops to MCC 0.40 (n=2) and 0.21 (n=4) on position-only disentanglement, versus 0.96-0.98 for the injective CNN†.
- On 3D shapes, the method reaches LD/MCC of 0.99 for position, color with n=2-4 objects, while SA-RP and SA-PC stay around 0.5-0.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use object-centric slot representations with the weakly supervised disentanglement loss and matching instead of a non-object-centric ResNet18 encoder | Causal representation learning on 2D Shapes · 2D Shapes test, fully sparse perturbations, multiple object counts and property sets | MCC | Helped |
| Use object-centric slot representations with the weakly supervised disentanglement loss and matching instead of a non-object-centric ResNet18 encoder | Causal representation learning on 2D Shapes · 2D Shapes test, fully sparse perturbations, multiple object counts and property sets | LD | Helped |
| Use object-centric slot representations with the weakly supervised disentanglement loss and matching instead of reconstruction-trained slot projections | SA-MESH · 3D Shapes test, fully sparse perturbations, position/color/size/rotation targets | LD | Helped |
| Use object-centric slot representations with the weakly supervised disentanglement loss and matching instead of reconstruction-trained slot projections | SA-MESH · 3D Shapes test, fully sparse perturbations, position/color/size/rotation targets | MCC | Helped |
| Represent shared object properties with a shared object-wise encoder rather than a fixed-size vector encoder | Causal representation learning with k objects · Synthetic multi-object settings covered by the theorem | required perturbations: k times fewer perturbations | Helped |
Surprising findings
- A non-injective ResNet18 fails badly once object order is dropped, with MCC falling toward 1/k as predicted.
- Even principal components of a vanilla reconstruction-trained SA-MESH give fairly high MCC on position tasks, attributed to positional encoding.
Limitations
- Experiments use only simple synthetic 2D and 3D scenes with homogeneous backgrounds and few objects (2-4).
- Each SA-MESH training takes about 12 hours on a single A100, and baselines were trained only once.
- Assumes fully shared properties across objects and 1-sparse perturbations with known structure in the main method.
Experimental setup
Datasets: synthetic 2D shapes dataset synthetic 3D shapes dataset
Benchmarks: 2D shapes (synthetic) 3D shapes (synthetic)
Metrics: MCC linear disentanglement R^2
Baselines: ResNet18 CNN (injective and non-injective variants) SA-RP (random projection on slots) SA-PC (principal components on slots) SA-LR (supervised linear regression upper bound)
Builds on: Slot Attention SA-MESH sparse perturbation disentanglement (Ahuja et al.) nonlinear ICA
Closest papers by meaning
- Multi-View Causal Representation Learning with Partial Observability (ICLR 2024 · similarity 0.89)
- A Sparsity Principle for Partially Observable Causal Representation Learning (ICML 2024 · similarity 0.89)
- Disentangled Representation Learning in Non-Markovian Causal Systems (NeurIPS 2024 · similarity 0.88)
- Causal Representation Learning Made Identifiable by Grouping of Observational Variables (ICML 2024 · similarity 0.87)
- Learning to Compose: Improving Object Centric Learning by Injecting Compositionality (ICLR 2024 · similarity 0.87)
- Temporally Consistent Object-Centric Learning by Contrasting Slots (CVPR 2025 · similarity 0.87)
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.