UCI Letter Recognition dataset

What the dataset is

The Letter Recognition dataset is one of the datasets distributed through the UCI Machine Learning Repository. Its objective is to identify each of a large number of black-and-white rectangular pixel displays as one of the 26 capital letters in the English alphabet^[1]. The repository lists 20,000 instances and 16 features and classifies the task as multivariate classification^[1] — that is, every example is described by several numbers at once and must be sorted into one of many categories rather than into one of two.

The repository records the dataset as having been donated on 12/31/1990^[1].

How the images were made

The character images were based on 20 different fonts, and each letter was randomly distorted to produce 20,000 unique stimuli^[1]. The point of the distortion was to stop the task from becoming a matter of matching a handful of fixed typefaces: if the same twenty shapes simply repeated, a learner could memorise them. By varying the shape of every instance, the dataset forces a model to generalise across different renderings of the same letter.

Each stimulus was then converted into 16 primitive numerical attributes, statistical moments and edge counts, scaled into integer values from 0 through 15^[1]. A statistical moment summarises the distribution of pixel values across the image — roughly, how dark the character is and how that darkness is spread out — while edge counts record how often sharp transitions between light and dark occur, which is a compact way of describing how much outline and internal structure a shape has. Scaling every attribute into the same small integer range keeps the 16 measurements on a common footing, so that no single one dominates simply because of the units it was recorded in.

How it is used

The repository notes that researchers typically train on the first 16,000 items and predict the letter category for the remaining 4,000^[1]. This is a convention rather than a rule of the data: the first block of the file acts as the training set and the last block as a held-out test set, so that results reported by different people can be compared on the same split. Because the correct answer for each of the 26 classes is known, performance is easy to score — a model's error rate is simply the share of held-out items assigned the wrong letter, and the 26-class structure makes the task considerably harder than a yes-or-no decision.

Where it sits in the archive

The dataset belongs to the UCI Machine Learning Repository, which the repository's About page says was created as an ftp archive in 1987 by UCI PhD student David Aha^[2]. The About page describes the repository as a collection of databases, domain theories and data generators used for the empirical analysis of machine learning algorithms^[2]. Letter Recognition dates from the archive's early period, when a shared collection of standard problems was what allowed researchers in different places to compare methods at all. Its format reflects that purpose: a plain table of integers with a known target column, small enough to download and run on the hardware of the time.

Limits

Sixteen attributes are a heavy compression of a pixel image. Details of a letter's shape that do not show up in the moments and edge counts are lost before any algorithm sees the data, so performance on this dataset measures how well a method uses a fixed, coarse summary rather than how well it reads characters. The task is also a cleaned-up version of reading: single characters, isolated on a plain background, with no words, spacing or context to help. Those constraints are what made it a useful benchmark, and they are also why results on it say less about recognising text in the world than the name might suggest.

^[1]: UCI Machine Learning Repository, “Letter Recognition dataset”.

^[2]: UCI Machine Learning Repository, “About the UCI Machine Learning Repository”.