Disentangling Symbols and Movements
Factor-VAE on NAR Dataset for few-shot visual analogical reasoning
Applied Factor-VAE to disentangle complex primitive transformations in a few-shot visual analogical reasoning task using the NAR (Neural Analogical Reasoning) dataset. Demonstrated interpretability of the learned factors through latent traversal analysis, showing that the model can separate symbolic structure from visual movement patterns.
Key contributions:
- Adapted Factor-VAE for disentanglement of visual analogy primitives
- Latent traversal analysis to verify factor interpretability
- Evaluation in a few-shot visual analogical reasoning setting