Para-CFlows: $C^k$-universal diffeomorphism approximators as superior neural surrogates

Authors: Junlong Lyu, Zhitang Chen, Chang Feng, Wenjing Cun, Shengyu Zhu, Yanhui Geng, ZHIJIE XU, Chen Yongwei

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

Reproducibility Variable Result LLM Response
Research Type Experimental In practice, we apply Para-CFlows as a neural surrogate model in contextual Bayesian optimization tasks, to demonstrate its superiority over other neural surrogate models in terms of optimization performance and gradient approximations. Code will be avaliable at https://gitee.com/mindspore/models/ tree/master/research/bo/paracflow.
Researcher Affiliation Industry Junlong Lyu Huawei Noah s Ark Lab Hong Kong SAR, China lyujunlong@huawei.com Zhitang Chen Huawei Noah s Ark Lab Hong Kong SAR, China chenzhitang2@huawei.com Chang Feng Huawei Noah s Ark Lab China cunwenjing@huawei.com Wenjing Cun Huawei Noah s Ark Lab China fengchang1@huawei.com Shengyu Zhu Huawei Noah s Ark Lab China zhushengyu@huawei.com Yanhui Geng Huawei Noah s Ark Lab Hong Kong SAR, China geng.yanhui@huawei.com Zhijie Xu Huawei China xuzhijie@huawei.com Yongwei Chen Huawei China chenyongwei@huawei.com
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks. It provides a network structure diagram but no code-like procedures.
Open Source Code Yes Code will be avaliable at https://gitee.com/mindspore/models/ tree/master/research/bo/paracflow.
Open Datasets Yes We generate samples {(xi, yi)}1 i 30000 with xi U[ 1, 1]2 and yi U[0, 1] independently. Under the polar coordinate representation, i.e., xi = (ρi cos θi, ρi sin θi), we calculate the target fyi(xi) = (ρi cos θi, ρi sin θi) according to Eq. (3). We use a 6-layers affine coupling flows, each layer composed with a random permute. The coupling functions σ, t are implemented by 1-hidden-layer Multi-Layer Perceptron (MLP) with hidden-unit number comparable to input dimension. We pad zeros to the input x, and use the first two output dimensions to compute the loss. [...] We ensemble Para-CFlows as a novel surrogate model for BO with three well-known benchmark functions: Rastrigin, Ackley and Trid [41].
Dataset Splits No The paper mentions generating training samples and using benchmark functions, but it does not specify explicit train/validation/test splits, proportions, or sample counts for reproducibility of the data partitioning methodology.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper describes the architecture of the neural networks used (e.g., 6-layer affine coupling flows, 1-hidden-layer MLPs) but does not list specific software dependencies with version numbers (e.g., Python version, specific deep learning framework like PyTorch or TensorFlow with versions, CUDA version).
Experiment Setup No The paper provides some architectural details like "6-layers affine coupling flows" and "1-hidden-layer Multi-Layer Perceptron (MLP) with hidden-unit number comparable to input dimension." However, it lacks specific hyperparameter values (e.g., learning rate, batch size, number of epochs, optimizer settings) which are crucial for full reproducibility.