基于区块链和智能合约的领域知识规约智能组件与算法

"本数据集为基于区块链和智能合约的新型去中心化人工智能协作框架（称为学习市场，Learning Markets）在MNIST数据集上的实验数据。MNIST数据集是被广泛使用的图像分类开源数据集，包含60000张训练样本和10000张测试样本，每个数据样本都是28×28像素的0-9之间的手写数字。我们考虑数据公平划分和数据不公平划分两种测试场景。 1）数据公平划分。组织者将完整数据集以1：4：65的比例随机划分为分发给训练者的训练集，分发给验证者的验证集和不分发的本地数据集三部分。训练者分别提交在训练集上得到的CNN，LSTM和BiLSTM模型。2）数据不公平划分。令前三个训练者（T1, T2, T3）分别持有任意5个手写数字的各2000个样本（共10000个样本），第四个训练者（T4）持有全部手写数字的各2000个样本（共20000个样本），并分别从各训练集中以0.1的比例随机抽样构成验证集（共5000个样本）。训练者分别提交在训练集上得到的CNN模型。该数据集包括选取的MNIST实验数据集，神经网络模型文件，智能合约源码和执行整个实验过程的测试代码。数据量共计1.22GB。"This dataset comprises experimental data derived from tests conducted on the MNIST dataset using a novel decentralized artificial intelligence (AI) collaboration framework based on blockchain and smart contracts, termed Learning Markets. The MNIST dataset is a widely adopted open-source image classification dataset, consisting of 60,000 training samples and 10,000 test samples. Each sample is a handwritten digit between 0 and 9 with a resolution of 28×28 pixels. Two test scenarios are considered: fair data partitioning and unfair data partitioning. 1) Fair data partitioning. The organizer randomly splits the full dataset into three parts at a ratio of 1:4:65: the training set distributed to trainers, the validation set distributed to validators, and the local dataset that is not distributed. Trainers separately submit the CNN, LSTM, and BiLSTM models trained on their respective training sets. 2) Unfair data partitioning. The first three trainers (T1, T2, T3) each hold 2,000 samples for any 5 handwritten digits, totaling 10,000 samples overall. The fourth trainer (T4) holds 2,000 samples for all 10 handwritten digits, totaling 20,000 samples overall. A validation set with 5,000 samples in total is randomly sampled from each training set at a ratio of 0.1. Trainers separately submit the CNN models trained on their respective training sets. This dataset includes the selected MNIST experimental dataset, neural network model files, smart contract source code, and test code for executing the entire experimental workflow. The total data volume is 1.22 GB.

搜集汇总

背景与挑战

背景概述

本数据集提供了基于区块链和智能合约的去中心化人工智能协作框架（学习市场）在MNIST手写数字数据集上的实验数据。它包含数据公平划分和不公平划分两种测试场景，涉及CNN、LSTM和BiLSTM等模型，并集成了模型文件、智能合约源码及测试代码。数据总量为1.23GB，共72个文件，由中国科学院自动化研究所发布。