Improving Neural Logic Machines via Failure Reflection
Authors: Zhiming Li, Yushi Cao, Yan Zheng, Xu Liu, Bozhi Wu, Tianlin Li, Xiufeng Xu, Junzhe Jiang, Yon Shin Teo, Shang-Wei Lin, Yang Liu
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on multiple relational reasoning and decision-making tasks demonstrate the effectiveness of FRGR in improving performance, generalization, training efficiency, and data efficiency. |
| Researcher Affiliation | Collaboration | 1Nanyang Technological University, Singapore 2Tianjin university, Tianjin, China 3National University of Singapore, Singapore 4Hong Kong Polytechnic University, Hong Kong 5Continental Automotive Singapore Pte. Ltd., Singapore. |
| Pseudocode | Yes | Algorithm 1 FRGR framework. Input: maximum iteration step T, NLMs model fθ parameterized by θ, regulatory coefficient β" and "Algorithm 1 FRGR framework for reinforcement learning tasks. Input: maximum number of episodes Epi, maximum iteration step T, NLMs model fθ parameterized by θ, regulatory coefficient β |
| Open Source Code | Yes | Our code is available at https://sites.google.com/ view/frgr-icml24. |
| Open Datasets | Yes | We follow previous work (Dong et al., 2018; Zimmer et al., 2023) and evaluate our framework on two reasoning benchmarks: relational reasoning and reinforcement learning: Relational reasoning. The relational reasoning tasks contain two major categories: Family Tree Reasoning (Dong et al., 2018; Evans & Grefenstette, 2018) and General Graph Reasoning (Graves et al., 2016; Dong et al., 2018; Zimmer et al., 2023). |
| Dataset Splits | Yes | For the Family Tree Reasoning task, all the models are trained on family trees with 20 family members and tested on samples of family sizes of 20 (performance) and 100 (generalization)." and "Epochs measures the number of training epochs required to reach optimal success rate on the validation set. |
| Hardware Specification | Yes | We conduct all experiments on a Ubuntu 18.05 server with 48 cores of Intel Xeon Siver 4214 CPU, 4 NVIDIA Quadro RTX 8000 GPUs, and 252GB RAM. |
| Software Dependencies | No | The paper mentions 'Ubuntu 18.05 server' and algorithms like 'Adam optimizer' and 'REINFORCE algorithm', but does not list specific software libraries or frameworks with version numbers (e.g., PyTorch, TensorFlow, CUDA versions) used for implementation. |
| Experiment Setup | Yes | For our method, we strictly follow the training settings of both NLM (Dong et al., 2018) and DLM(Zimmer et al., 2023). They are trained using Adam optimizer (Kingma & Ba, 2014) with a 0.005 learning rate. For all the relational reasoning tasks, the Softmax Cross Entropy is used as the loss function. ...The batch size is set to be 4 across all the experiments. The regulatory coefficient β is set to be 0.1 and τ is set to be 100 for all tasks. |