Combining Symbolic Expressions and Black-box Function Evaluations in Neural Programs

Authors: Forough Arabshahi, Sameer Singh, Animashree Anandkumar

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We present an evaluation benchmark for this task to demonstrate our proposed model combines symbolic reasoning and function evaluation in a fruitful manner, obtaining high accuracies in our experiments. 4 EXPERIMENTS AND RESULTS
Researcher Affiliation Academia Forough Arabshahi University of California Irvine, CA farabsha@uci.edu Sameer Singh University of California Irvine, CA sameer@uci.edu Animashree Anandkumar California Institute of Technology Pasadena, CA anima@caltech.edu
Pseudocode No The paper describes methods and processes but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Our dataset generation method, proposed model, and data is available here: https://github.com/ Forough A/neural Math
Open Datasets Yes Our dataset generation method, proposed model, and data is available here: https://github.com/ Forough A/neural Math
Dataset Splits No The paper states: 'In this experiment we randomly split all of the generated data that includes equations of depths 1 to 4 into train and test partitions with an 80%/20% split ratio.' It specifies train and test splits, but does not explicitly mention a separate validation set split or its proportion.
Hardware Specification No The paper mentions thanking Amazon Inc. for AWS credits, implying cloud usage, but does not provide specific hardware details such as GPU/CPU models or instance types used for experiments.
Software Dependencies No The paper mentions software like Mx Net, Sympy, and Adam optimizer, but does not provide specific version numbers for any of them.
Experiment Setup Yes We use L2-regularization as well as dropout to avoid overfitting, and train all the models for 100 epochs. We have tuned for the hidden dimension {10,20,50}, the optimizers {SGD, NAG (Nesterov accelerated SGD), RMSProp, Adam, Ada Grad, Ada Delta, DCASGD, SGLD (Stochastic Gradient Riemannian Langevin Dynamics)}, dropout rate {0.2,0.3}, learning rate {10 3, 10 5}, regularization ratio {10 4, 10 5} and momentum {0.2,0.7}. Most of the networks achieved their best performance using Adam optimizer Kingma & Ba (2014) with a learning rate of 0.001 and a regularization ratio of 10 5. Hidden dimension and dropout varies under each of the scenarios.