A collection of examples gathered to teach or test an AI.
A dataset is a set of examples put together for an AI: photos with labels, sentences with translations, or huge piles of text collected from the web. It's usually split into a part for training and a part kept back for testing, so the model is checked on questions it hasn't seen.
An AI can only learn what's in its data. If the data leans one way, the answers lean too; if it contains mistakes, the model learns the mistakes. So what was collected, where from, and with whose permission matter as much as how much. ImageNet, released in 2009, is often cited as proof that a dataset can lift a whole field.
Think of it like this
A dataset is an AI's workbook. If the workbook only has problems from one chapter, you only get good at that chapter. And if the test questions are already in the workbook, the test measures memory, not skill.
For example
MNIST, a set of 70,000 photos of handwritten digits, is one of the most famous practice datasets.
Records about this
Source-checked moments from AI history.
The diversity-curation paradigm. At the same time, it passed copyright burdens such as Books3 on to later models
Released a large-scale image dataset based on the WordNet hierarchy, creating the recognition that data is the bottleneck
Established a practical CNN architecture combining convolution, pooling and fully connected layers, and made MNIST the standard testbed