Learn to Explain Efficiently via Neural Logic Inductive Learning
Authors: Yuan Yang, Le Song
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments, compared with the state-of-the-art methods, we find NLIL can search for rules that are x10 times longer while remaining x3 times faster. We also show that NLIL can scale to large image datasets, i.e. Visual Genome, with 1M entities. |
| Researcher Affiliation | Academia | Yuan Yang & Le Song Georgia Institute of Technology yyang754@gatech.edu, lsong@cc.gatech.edu |
| Pseudocode | No | The paper describes the computational process and architecture, but it does not contain any formally labeled "Pseudocode" or "Algorithm" blocks with structured steps. |
| Open Source Code | Yes | Our implementation is available at https://github.com/gblackout/NLIL. |
| Open Datasets | Yes | Benchmark datasets: (i) Even-and-Successor (ES) benchmark is introduced in (Evans & Grefenstette, 2018)... (ii) FB15K-237 is a subset of the Freebase knowledge base (Toutanova & Chen, 2015)... (iii) WN18 (Bordes et al., 2013) is the subset of Word Net... We conduct experiments on Visual Genome dataset (Krishna et al., 2016). The original dataset is highly noisy (Zellers et al., 2018), so we use a pre-processed version available as the GQA dataset (Hudson & Manning, 2019). |
| Dataset Splits | No | The benchmark datasets are split into train/valid/test sets. For each query x, Pk, x , the model generates a ranking list over all possible groundings of predicate Pk, with other groundtruth triplets filtered out. Then MRR is the average of the reciprocal rank of the queries in their corresponding lists, and Hits@10 is the percentage of queries that are ranked within the top 10 in the list. |
| Hardware Specification | Yes | All experiments are conducted on a machine with i7-8700K, 32G RAM and one GTX1080ti. |
| Software Dependencies | No | The paper provides a link to its implementation but does not list specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions) within the text. |
| Experiment Setup | Yes | For NLIL, we create separate Transformer blocks for each target predicate. ... We use the embedding size d = 32. We use 3 layers of multi-head attentions for each Transformer network. The number of attention heads are set to number of heads = 4 for encoder, and the first two layers of the decoder. The last layer of the decoder has one attention head to produce the final attention required for rule evaluation. ... For KB completion task, we set the number of operator calls T = 2 and formula combinations L = 0... For VG dataset, we set T = 3, L = 2 and C = 4. |