HDK: Toward High-Performance Deep-Learning-Based Kirchhoff Analysis
Authors: Xinying Wang, Olamide Timothy Tawose, Feng Yan, Dongfang Zhao997-1004
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We implement the proposed approach with the latest machine learning framework and the parallel computing library, and evaluate it with real-world engineering applications showing promising results: the accuracy is as high as 99.6% with up to 8 performance improvement over the state-of-the-art. |
| Researcher Affiliation | Academia | Xinying Wang, Olamide Timothy Tawose, Feng Yan, Dongfang Zhao University of Nevada Reno, NV 89557, USA {xinyingw, otawose}@nevada.unr.edu, {fyan, dzhao}@unr.edu |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We have implemented the proposed method, mainly with Python and MPI, and released the source code hosted at Github: https://github.com/hpdic/ HDK On Electrode Arrays. |
| Open Datasets | No | Both data sets are collected from our wet lab with measured Zs using 64 64 electrode arrays. The core data set comprises 100 data points, which are augmented into 1,400 data points through up-to1% random errors.The paper describes how the dataset was generated/collected but does not provide concrete access information (link, DOI, formal citation) for public availability. |
| Dataset Splits | Yes | For both data sets, we hold out 20% of the data for prediction, and the remaining 80% for training. We use 10-fold cross-validation and report the variance in the error bar if it is noticeable. |
| Hardware Specification | Yes | Our experiments are primarily carried out on the Amazon Web Services (AWS). Table 1 lists the instance types for the evaluation. For the parallel forward labelling experiment, we use both of the t2.2xlarge and the c5.18xlarge instances. For model training, we use only the t2.2xlarge instance. Table 1: AWS instances used for evaluation. Instance CPU Memory t2.2xlarge Intel Xeon E5-2686 32 GB c5.18xlarge Intel Xeon Platinum 8124M 144 GB |
| Software Dependencies | Yes | All instances are installed with Ubuntu 18.04, Anaconda 2019.3, Python 3.7, Num Py 1.15.4, Sci Py 0.17.0, mpi4py v2.0.0, and mpich2 v1.4.1. |
| Experiment Setup | Yes | DL Models The models are trained through Keras (Keras 2019), a high-level API built upon Tensor Flow (Tensor Flow 2019). Models in Keras are defined as a sequence of layers. We create a sequential model and add layers one at a time until the network reaches high accuracy. The number of inputs are initially set to input dim = 512 for both models. In both models, we use a fully-connected network structure with two hidden layers, and the number of neurons is 2,048. Table 2 illustrates the network architecture of the DL models. Table 2: DL network architecture for 64 64 arrays. Layer Input Dense-1 Dense-2 Output Neuron 100 1,400 2,048 2,048 100 1,400. During the training, we initialize the network weights with a custom initializer (He initialization) as we use the rectifier relu activation function on the first two layers for a reduced likelihood of vanishing gradient. We apply no activation function to the output layer because we aim to train a regression model instead of a classification model. When compiling, we use the mean squared error loss function and the gradient descent algorithm adam that appears to be highly efficient for our data. |