Symbolic Learning to Optimize: Towards Interpretability and Scalability

Authors: Wenqing Zheng, Tianlong Chen, Ting-Kuei Hu, Zhangyang Wang

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The evaluation results are reported in Fig.2 and Table 4.
Researcher Affiliation Academia 1The University of Texas at Austin, 2Texas A&M University
Pseudocode Yes Figure 7: The pseudo-code for the core mutation operation used in the SR algorithm.
Open Source Code Yes Codes are available at: https: //github.com/VITA-Group/Symbolic-Learning-To-Optimize.
Open Datasets Yes The L2O meta-training was performed on multiple randomly-initialized Le Nets on the MNIST dataset. ... Res Net-50, Res Net-152 and Mobile Net V2, on the Cifar-10 and Cifar-100 datasets.
Dataset Splits No No explicit percentages or sample counts for training, validation, and test splits of the datasets (MNIST, CIFAR-10, CIFAR-100) were found in the paper's main text. Although a 'validation set' is mentioned for symbolic distillation, the overall dataset splits for model training are not specified.
Hardware Specification Yes The execution times for SR and computing TPF/MC are measured on the 2.6 GHz Intel Core i7 CPU with 16 GB 2400 MHz DDR4 Memory, and other other computation times are measured on the Nvidia A6000 GPU.
Software Dependencies No No specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow, specific libraries or solvers with versions) were explicitly stated in the paper.
Experiment Setup Yes We used 128 batch size, and in both meta-fine-tune phase and final evaluation phase, we trained the CNN optimizees for 200 epochs. ... The number of iterations: ...300... The population number: ...200... Dataset size: ...5000 samples... For both optimizers, we used lr = 0.0005. For Adam W, we used β1 = 0.9, β2 = 0.999.