Neural Program Lattices
Authors: Chengtao Li, Daniel Tarlow, Alexander L. Gaunt, Marc Brockschmidt, Nate Kushman
ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the capability of NPL to learn to perform long-hand addition and arrange blocks in a grid-world environment. Experiments show that it performs on par with NPI while using weak supervision in place of most of the strong supervision, thus indicating its ability to infer the high-level program structure from examples containing only the low-level operations. In this section, we demonstrate the capability of NPL to learn on both the long-hand addition task (ADDITION) from Reed & de Freitas (2016) and a newly introduced task involving arranging blocks in a grid-world (NANOCRAFT). |
| Researcher Affiliation | Collaboration | Chengtao Li Massachusetts Institute of Technology Cambridge, MA 02139, USA ctli@mit.edu Daniel Tarlow, Alexander L. Gaunt, Marc Brockschmidt, Nate Kushman Microsoft Research Cambridge, CB1 2FB, UK {dtarlow,algaunt,mabrocks,nkushman}@microsoft.com |
| Pseudocode | No | The paper does not contain any explicitly labeled pseudocode or algorithm blocks. Procedural descriptions and mathematical formulations are integrated into the text. |
| Open Source Code | No | The paper does not contain any statements or links indicating that the source code for the described methodology is publicly available. |
| Open Datasets | No | The paper describes tasks (ADDITION, NANOCRAFT) and how data was generated for them, or references tasks from other papers, but does not provide a direct link, DOI, or explicit statement of public availability for the specific datasets used in their experiments. For example, for ADDITION it states, 'The programs and elementary operations for ADDITION are identical to those in Reed & de Freitas (2016)', indicating the task definition, not a readily accessible dataset from that paper. |
| Dataset Splits | Yes | All experiments were run with 10 different random seeds, and the best model was chosen using a separate validation set which is one-quarter the size of the training set. We ran all models using 100 different seeds for each model. We then sampled 25 seed subsets, with replacement. For each subset, we choose the best seed using a validation set which was one-quarter the size of the original dataset, but consisted only of 10-digit samples. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used to run its experiments. |
| Software Dependencies | No | The paper mentions training with ADAM and using LSTM cells, but it does not specify concrete software components with version numbers (e.g., Python, PyTorch, TensorFlow versions, or specific library versions). |
| Experiment Setup | Yes | For all tasks, we train the NPL using ADAM (Kingma & Ba, 2015) with base learning rate of 10 4 and batch size of 1. We decay the learning rate by a factor of 0.95 every 10,000 iterations. The size of the hidden states is set to 128 for both ADDITION and NANOCRAFT. |