Learning Discrete Concepts in Latent Hierarchical Models
Authors: Lingjing Kong, Guangyi Chen, Biwei Huang, Eric Xing, Yuejie Chi, Kun Zhang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We substantiate our theoretical claims with synthetic data experiments. Further, we discuss our theory s implications for understanding the underlying mechanisms of latent diffusion models and provide corresponding empirical evidence for our theoretical insights. |
| Researcher Affiliation | Academia | 1Carnegie Mellon University 2Mohamed bin Zayed University of Artificial Intelligence 3University of California San Diego |
| Pseudocode | Yes | Algorithm 1: The overall procedure for Rank-based Discrete Latent Causal Model Discovery. |
| Open Source Code | Yes | The code can be found here. |
| Open Datasets | No | We generate the hierarchical model G with randomly sampled parameters, and follow [24] to build the generating process from d to the observed variables x (i.e., graph Γ) by a Gaussian mixture model. |
| Dataset Splits | No | The paper describes generating synthetic data and evaluating on specific graphs, but does not specify training, validation, and test dataset splits. |
| Hardware Specification | Yes | We conduct our experiments on a cluster of 64 CPUs. All experiments can be finished within half an hour. The search algorithm implementation is adapted from Dong et al. [20]. We conduct all our experiments on 2 Nvidia L40 GPUs. |
| Software Dependencies | No | We employ the pre-trained latent diffusion model [28] SD v1.4 across all our experiments. |
| Experiment Setup | Yes | Experimental setup. We generate the hierarchical model G with randomly sampled parameters, and follow [24] to build the generating process from d to the observed variables x (i.e., graph Γ) by a Gaussian mixture model. The inference process consists of 50 steps. For experiments in Section 7.1, we inject concepts by appending keywords to the original prompt. We evaluate ranks in {2, 4, 8} and scales {1, 2, 3, 4, 5}. |