n-digit MNIST

MNIST 手写数字可以说是机器学习研究中最受欢迎的数据集。尽管最先进的学习模型很久以前就可能在该基准上达到了可实现的最佳性能，但数据集本身对研究界仍然有用，为新方法提供了一个简单的健全性检查：如果它不起作用MNIST，它在任何地方都不起作用！
我们在这里介绍 MNIST 的 n 位变体。通过为每个数据点添加更多数字，可以成倍地增加数据集的类别数量。尽管如此，它们仍然利用数据的简单性和轻量级特性。这些数据集提供了一个简单而有用的玩具示例，例如人脸嵌入。人们还可以在单个数字和例如数字之间进行类比。面属性。在这种情况下，数据集用于提供对嵌入算法的快速洞察，以便将其扩展到更现实、训练缓慢的问题。
由于潜在的专有性问题和更大的灵活性，我们发布了用于从原始 MNIST 数据集生成数据集的代码，而不是发布图像本身。出于基准测试的目的，我们发布了四个标准数据集，这些数据集再次通过代码生成，但具有确定性。

The MNIST handwritten digit dataset is widely regarded as one of the most popular datasets in machine learning research. Although state-of-the-art machine learning models have long achieved the best achievable performance on this benchmark, the dataset itself remains valuable to the research community, serving as a straightforward sanity check for new methods: if a new method fails on MNIST, it will fail anywhere else!

We hereby introduce the n-digit variant of MNIST. By incorporating additional digits into each data point, the number of categories in the dataset can be exponentially increased. Nevertheless, these variants still retain the simplicity and lightweight nature of the original MNIST data. These datasets offer simple yet useful toy examples, such as facial embedding tasks. One can also draw analogies between individual handwritten digits and other targets such as facial attributes. In such scenarios, these datasets can provide rapid insights into embedding algorithms, facilitating their extension to more realistic, time-consuming training problems.

Due to potential proprietary concerns and greater flexibility, we release the code for generating datasets from the original MNIST corpus rather than distributing the raw image files directly. For benchmarking purposes, we have released four standard datasets, which are again generated using the provided code but with deterministic generation workflows.

搜集汇总

背景与挑战

背景概述

n-digit MNIST 是基于 MNIST 手写数字数据集构建的变体，通过增加每个数据点的数字数量来扩展类别，为嵌入算法研究提供轻量级的玩具示例。该数据集由 Google 于 2018 年发布，主要用于机器学习方法的快速健全性测试，并提供了生成代码而非直接图像。