Differentiable Synthesis of Program Architectures
Authors: Guofeng Cui, He Zhu
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiment results on four sequence classification tasks demonstrate that our program synthesizer excels in discovering program architectures that lead to differentiable programs with higher F1 scores, while being more efficient than state-of-the-art program synthesis methods. |
| Researcher Affiliation | Academia | Guofeng Cui Department of Computer Science Rutgers University gc669@cs.rutgers.edu, He Zhu Department of Computer Science Rutgers University hz375@cs.rutgers.edu |
| Pseudocode | Yes | Our differentiable program architecture synthesis method is outlined in (the first while loop of) Algorithm 1. |
| Open Source Code | Yes | We have implemented Algorithm 1 in a tool named d Pads (domain-specific Program architecture differentiable synthesis) [19] |
| Open Datasets | Yes | Crim13 Dataset. The dataset collects social behaviors of a pair of mice... [20]; Fly-vs-fly Dataset. We use the Boy-meets-boy, Aggression and Courtship datasets... [21]. URL https://data.caltech.edu/records/ 1893.; Basketball Dataset. The dataset tracks the movement of a basketball... [22]; Skeletics 152 Dataset. The dataset [23] contains 152 human pose actions as well as related You Tube Videos subsampled from Kinetics-700 [24]. |
| Dataset Splits | Yes | We partition a dataset to training, validation, and test datasets. ... In total we have 12404, 3077, and 2953 trajectories in the training set, validation set, and test set respectively. (for Crim13) |
| Hardware Specification | Yes | All experiments were performed on Intel 2.3-GHz Xeon CPU with 16 cores, equipped with an NVIDIA Quadro RTX 6000 GPU. |
| Software Dependencies | No | The paper mentions using the Adam optimizer [25] but does not provide specific version numbers for any software libraries, frameworks, or programming languages. |
| Experiment Setup | Yes | In Algorithm 1, we set N = 2 for top-N preservation and set graph expansion depth ds to 2. For evaluation, we compare d Pads with the state-of-the-art program learning algorithms A -NEAR and IDS-BB-NEAR [7]. More experiment settings including learning rates and training epochs are given in Appendix C.1 and C.2. |