Yann LeCun, Courant Institute, NYU

Corinna Cortes, Google Labs, New York

Christopher J.C. Burges, Microsoft Research, Redmond

The MNIST database of handwritten digits, available from this page, has a training set of 60,000 examples, and a test set
of 10,000 examples. It is a subset of a larger set available from NIST. The digits have been size-normalized and centered
in a fixed-size image.

It is a good database for people who want to try learning techniques and pattern recognition methods on real-world data
while spending minimal efforts on preprocessing and formatting.

Four files are available on this site:

train-images-idx3-ubyte.gz: training set images (9912422 bytes)

train-labels-idx1-ubyte.gz: training set labels (28881 bytes)
t10k-images-idx3-ubyte.gz: test set images (1648877 bytes)
t10k-labels-idx1-ubyte.gz: test set labels (4542 bytes)

please note that your browser may uncompress these files without telling you. If the files you downloaded have a
larger size than the above, they have been uncompressed by your browser. Simply rename them to remove the .gz
extension. Some people have asked me "my application can't open your image files". These files are not in any standard
image format. You have to write your own (very simple) program to read them. The file format is described at the bottom
of this page.

The original black and white (bilevel) images from NIST were size normalized to fit in a 20x20 pixel box while
preserving their aspect ratio. The resulting images contain grey levels as a result of the anti-aliasing technique used by the
normalization algorithm. the images were centered in a 28x28 image by computing the center of mass of the pixels, and
translating the image so as to position this point at the center of the 28x28 field.

With some classification methods (particuarly template-based methods, such as SVM and K-nearest neighbors), the error
rate improves when the digits are centered by bounding box rather than center of mass. If you do this kind of pre-
processing, you should report it in your publications.

The MNIST database was constructed from NIST's Special Database 3 and Special Database 1 which contain binary
images of handwritten digits. NIST originally designated SD-3 as their training set and SD-1 as their test set. However,
SD-3 is much cleaner and easier to recognize than SD-1. The reason for this can be found on the fact that SD-3 was
collected among Census Bureau employees, while SD-1 was collected among high-school students. Drawing sensible
conclusions from learning experiments requires that the result be independent of the choice of training set and test among
the complete set of samples. Therefore it was necessary to build a new database by mixing NIST's datasets.

The MNIST training set is composed of 30,000 patterns from SD-3 and 30,000 patterns from SD-1. Our test set was
composed of 5,000 patterns from SD-3 and 5,000 patterns from SD-1. The 60,000 pattern training set contained examples
from approximately 250 writers. We made sure that the sets of writers of the training set and test set were disjoint.

SD-1 contains 58,527 digit images written by 500 different writers. In contrast to SD-3, where blocks of data from each
writer appeared in sequence, the data in SD-1 is scrambled. Writer identities for SD-1 is available and we used this
information to unscramble the writers. We then split SD-1 in two: characters written by the first 250 writers went into our
new training set. The remaining 250 writers were placed in our test set. Thus we had two sets with nearly 30,000
examples each. The new training set was completed with enough examples from SD-3, starting at pattern # 0, to make a
full set of 60,000 training patterns. Similarly, the new test set was completed with SD-3 examples starting at pattern #
35,000 to make a full set with 60,000 test patterns. Only a subset of 10,000 test images (5,000 from SD-1 and 5,000 from
SD-3) is available on this site. The full 60,000 sample training set is available.

Many methods have been tested with this training set and test set. Here are a few examples. Details about the methods are
given in an upcoming paper. Some of those experiments used a version of the database where the input images where
deskewed (by computing the principal axis of the shape that is closest to the vertical, and shifting the lines so as to make it
vertical). In some other experiments, the training set was augmented with artificially distorted versions of the original
training samples. The distortions are random combinations of shifts, scaling, skewing, and compression.

TEST
CLASSIFIER PREPROCESSING ERROR Reference
RATE (%)
Linear Classifiers
linear classifier (1-layer NN) none 12.0 LeCun et al. 1998

1 of 5
 linear classifier (1-layer NN) deskewing 8.4 LeCun et al. 1998
pairwise linear classifier deskewing 7.6 LeCun et al. 1998
K-Nearest Neighbors
K-nearest-neighbors, Euclidean (L2) none 5.0 LeCun et al. 1998
K-nearest-neighbors, Euclidean (L2) none 3.09 Kenneth Wilder, U. Chicago
K-nearest-neighbors, L3 none 2.83 Kenneth Wilder, U. Chicago
K-nearest-neighbors, Euclidean (L2) deskewing 2.4 LeCun et al. 1998
deskewing, noise removal,
K-nearest-neighbors, Euclidean (L2) 1.80 Kenneth Wilder, U. Chicago
blurring
deskewing, noise removal,
K-nearest-neighbors, L3 1.73 Kenneth Wilder, U. Chicago
blurring
deskewing, noise removal,
K-nearest-neighbors, L3 1.33 Kenneth Wilder, U. Chicago
blurring, 1 pixel shift
deskewing, noise removal,
K-nearest-neighbors, L3 1.22 Kenneth Wilder, U. Chicago
blurring, 2 pixel shift
Keysers et al. IEEE PAMI
K-NN with non-linear deformation (IDM) shiftable edges 0.54
2007
K-NN with non-linear deformation Keysers et al. IEEE PAMI
shiftable edges 0.52
(P2DHMDM) 2007
subsampling to 16x16
K-NN, Tangent Distance 1.1 LeCun et al. 1998
pixels
shape context feature Belongie et al. IEEE PAMI
K-NN, shape context matching 0.63
extraction 2002
Boosted Stumps
boosted stumps none 7.7 Kegl et al., ICML 2009
products of boosted stumps (3 terms) none 1.26 Kegl et al., ICML 2009
boosted trees (17 leaves) none 1.53 Kegl et al., ICML 2009
stumps on Haar features Haar features 1.02 Kegl et al., ICML 2009
product of stumps on Haar f. Haar features 0.87 Kegl et al., ICML 2009
Non-Linear Classifiers
40 PCA + quadratic classifier none 3.3 LeCun et al. 1998
1000 RBF + linear classifier none 3.6 LeCun et al. 1998
SVMs
SVM, Gaussian Kernel none 1.4
SVM deg 4 polynomial deskewing 1.1 LeCun et al. 1998
Reduced Set SVM deg 5 polynomial deskewing 1.0 LeCun et al. 1998
Virtual SVM deg-9 poly [distortions] none 0.8 LeCun et al. 1998
DeCoste and Scholkopf,
Virtual SVM, deg-9 poly, 1-pixel jittered none 0.68
MLJ 2002
DeCoste and Scholkopf,
Virtual SVM, deg-9 poly, 1-pixel jittered deskewing 0.68
MLJ 2002
DeCoste and Scholkopf,
Virtual SVM, deg-9 poly, 2-pixel jittered deskewing 0.56
MLJ 2002
Neural Nets
2-layer NN, 300 hidden units, mean square
none 4.7 LeCun et al. 1998
error
2-layer NN, 300 HU, MSE, [distortions] none 3.6 LeCun et al. 1998
2-layer NN, 300 HU deskewing 1.6 LeCun et al. 1998
2-layer NN, 1000 hidden units none 4.5 LeCun et al. 1998
2-layer NN, 1000 HU, [distortions] none 3.8 LeCun et al. 1998
3-layer NN, 300+100 hidden units none 3.05 LeCun et al. 1998
3-layer NN, 300+100 HU [distortions] none 2.5 LeCun et al. 1998

2 of 5
 3-layer NN, 500+150 hidden units none 2.95 LeCun et al. 1998
3-layer NN, 500+150 HU [distortions] none 2.45 LeCun et al. 1998
3-layer NN, 500+300 HU, softmax, cross
none 1.53 Hinton, unpublished, 2005
entropy, weight decay
2-layer NN, 800 HU, Cross-Entropy Loss none 1.6 Simard et al., ICDAR 2003
2-layer NN, 800 HU, cross-entropy [affine
none 1.1 Simard et al., ICDAR 2003
distortions]
2-layer NN, 800 HU, MSE [elastic distortions] none 0.9 Simard et al., ICDAR 2003
2-layer NN, 800 HU, cross-entropy [elastic
none 0.7 Simard et al., ICDAR 2003
distortions]
NN, 784-500-500-2000-30 + nearest neighbor, Salakhutdinov and Hinton,
none 1.0
RBM + NCA training [no distortions] AI-Stats 2007
Ciresan et al. Neural
6-layer NN 784-2500-2000-1500-1000-500-10
none 0.35 Computation 10, 2010 and
(on GPU) [elastic distortions]
arXiv 1003.0358, 2010
committee of 25 NN 784-800-10 [elastic width normalization,
0.39 Meier et al. ICDAR 2011
distortions] deslanting
deep convex net, unsup pre-training [no
none 0.83 Deng et al. Interspeech 2010
distortions]
Convolutional nets
subsampling to 16x16
Convolutional net LeNet-1 1.7 LeCun et al. 1998
pixels
Convolutional net LeNet-4 none 1.1 LeCun et al. 1998
Convolutional net LeNet-4 with K-NN instead
none 1.1 LeCun et al. 1998
of last layer
Convolutional net LeNet-4 with local learning
none 1.1 LeCun et al. 1998
instead of last layer
Convolutional net LeNet-5, [no distortions] none 0.95 LeCun et al. 1998
Convolutional net LeNet-5, [huge distortions] none 0.85 LeCun et al. 1998
Convolutional net LeNet-5, [distortions] none 0.8 LeCun et al. 1998
Convolutional net Boosted LeNet-4,
none 0.7 LeCun et al. 1998
[distortions]
Trainable feature extractor + SVMs [no Lauer et al., Pattern
none 0.83
distortions] Recognition 40-6, 2007
Trainable feature extractor + SVMs [elastic Lauer et al., Pattern
none 0.56
distortions] Recognition 40-6, 2007
Trainable feature extractor + SVMs [affine Lauer et al., Pattern
none 0.54
distortions] Recognition 40-6, 2007
unsupervised sparse features + SVM, [no Labusch et al., IEEE TNN
none 0.59
distortions] 2008
Convolutional net, cross-entropy [affine
none 0.6 Simard et al., ICDAR 2003
distortions]
Convolutional net, cross-entropy [elastic
none 0.4 Simard et al., ICDAR 2003
distortions]
large conv. net, random features [no
none 0.89 Ranzato et al., CVPR 2007
distortions]
large conv. net, unsup features [no distortions] none 0.62 Ranzato et al., CVPR 2007
large conv. net, unsup pretraining [no
none 0.60 Ranzato et al., NIPS 2006
distortions]
large conv. net, unsup pretraining [elastic
none 0.39 Ranzato et al., NIPS 2006
distortions]
large conv. net, unsup pretraining [no
none 0.53 Jarrett et al., ICCV 2009
distortions]
large/deep conv. net,
1-20-40-60-80-100-120-120-10 [elastic none 0.35 Ciresan et al. IJCAI 2011
distortions]

3 of 5
 committee of 7 conv. net, 1-20-P-40-P-150-10
width normalization 0.27 +-0.02 Ciresan et al. ICDAR 2011
[elastic distortions]
committee of 35 conv. net, 1-20-P-40-P-150-10
width normalization 0.23 Ciresan et al. CVPR 2012
[elastic distortions]

References

[LeCun et al., 1998a]

Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. "Gradient-based learning applied to document recognition."
Proceedings of the IEEE, 86(11):2278-2324, November 1998. [on-line version]

FILE FORMATS FOR THE MNIST DATABASE

The data is stored in a very simple file format designed for storing vectors and multidimensional matrices. General info on
this format is given at the end of this page, but you don't need to read that to use the data files.

All the integers in the files are stored in the MSB first (high endian) format used by most non-Intel processors. Users of
Intel processors and other low-endian machines must flip the bytes of the header.

There are 4 files:

train-images-idx3-ubyte: training set images

train-labels-idx1-ubyte: training set labels
t10k-images-idx3-ubyte: test set images
t10k-labels-idx1-ubyte: test set labels

The training set contains 60000 examples, and the test set 10000 examples.

The first 5000 examples of the test set are taken from the original NIST training set. The last 5000 are taken from the
original NIST test set. The first 5000 are cleaner and easier than the last 5000.

TRAINING SET LABEL FILE (train-labels-idx1-ubyte):

[offset] [type] [value] [description]
0000 32 bit integer 0x00000801(2049) magic number (MSB first)
0004 32 bit integer 60000 number of items
0008 unsigned byte ?? label
0009 unsigned byte ?? label
........
xxxx unsigned byte ?? label

The labels values are 0 to 9.

TRAINING SET IMAGE FILE (train-images-idx3-ubyte):

[offset] [type] [value] [description]
0000 32 bit integer 0x00000803(2051) magic number
0004 32 bit integer 60000 number of images
0008 32 bit integer 28 number of rows
0012 32 bit integer 28 number of columns
0016 unsigned byte ?? pixel
0017 unsigned byte ?? pixel
........
xxxx unsigned byte ?? pixel

Pixels are organized row-wise. Pixel values are 0 to 255. 0 means background (white), 255 means foreground (black).

TEST SET LABEL FILE (t10k-labels-idx1-ubyte):

[offset] [type] [value] [description]
0000 32 bit integer 0x00000801(2049) magic number (MSB first)
0004 32 bit integer 10000 number of items
0008 unsigned byte ?? label
0009 unsigned byte ?? label
........
xxxx unsigned byte ?? label

The labels values are 0 to 9.

TEST SET IMAGE FILE (t10k-images-idx3-ubyte):

[offset] [type] [value] [description]
0000 32 bit integer 0x00000803(2051) magic number
0004 32 bit integer 10000 number of images
0008 32 bit integer 28 number of rows
0012 32 bit integer 28 number of columns
0016 unsigned byte ?? pixel
0017 unsigned byte ?? pixel

4 of 5
 ........
xxxx unsigned byte ?? pixel

Pixels are organized row-wise. Pixel values are 0 to 255. 0 means background (white), 255 means foreground (black).

THE IDX FILE FORMAT

the IDX file format is a simple format for vectors and multidimensional matrices of various numerical types.

The basic format is

magic number
size in dimension 0
size in dimension 1
size in dimension 2
.....
size in dimension N
data

The magic number is an integer (MSB first). The first 2 bytes are always 0.

The third byte codes the type of the data:

0x08: unsigned byte
0x09: signed byte
0x0B: short (2 bytes)
0x0C: int (4 bytes)
0x0D: float (4 bytes)
0x0E: double (8 bytes)

The 4-th byte codes the number of dimensions of the vector/matrix: 1 for vectors, 2 for matrices....

The sizes in each dimension are 4-byte integers (MSB first, high endian, like in most non-Intel processors).

The data is stored like in a C array, i.e. the index in the last dimension changes the fastest.

Happy hacking.

The digit images in the MNIST set were originally selected and experimented with by Chris Burges and Corinna Cortes
using bounding-box normalization and centering. Yann LeCun's version which is provided on this page uses centering by
center of mass within in a larger window.

Yann LeCun, Professor

The Courant Institute of Mathematical Sciences
New York University

Corinna Cortes, Research Scientist

Google Labs, New York
corinna at google dot com

5 of 5