Results on various benchmark data sets1

Handwritten digits

NN architecture Dataset Distortions Test Error [%] Info
MLP:2500-2000-1500-1000-500-10 MNIST no 1.57 -
MLP:2000-1500-1000-500-10 MNIST 5% translation 0.94 -
MLP:2500-2000-1500-1000-500-10 MNIST affine + elastic 0.35 Deep Big Simple Neural Nets Excel on Handwritten Digit Recognition, NC, Dec. 2010
MLP committee:2500-2000-1500-1000-500-10 MNIST affine + elastic 0.31 Handwritten Digit Recognition with a Committee of Deep Neural Nets on GPUs
CNN 20M-40M-60M-80M-100M-120M-150N MNIST affine + elastic 0.35 High-Performance Neural Networks for Visual Object Classification, IJCAI 2011
committee of 7 CNNs MNIST affine + elastic 0.27±0.02 Convolutional Neural Network Committees For Handwritten Character Classification, ICDAR 2011
MCDNN MNIST affine + elastic 0.23 (first human competitive result) Multi-Column Deep Neural Networks For Image Classification, CVPR 2012

Handwritten characters (NIST SD 19)

NN architecture Dataset Defformations Test Error [%] Info
committee of 7 CNNs NIST digits affine + elastic 0.81±0.02 Convolutional Neural Network Committees For Handwritten Character Classifications, ICDAR 2011
committee of 7 CNNs NIST small letters affine + elastic 7.71±0.14 Convolutional Neural Network Committees For Handwritten Character Classifications, ICDAR 2011
committee of 7 CNNs NIST big letters affine + elastic 1.91±0.06 Convolutional Neural Network Committees For Handwritten Character Classifications, ICDAR 2011
committee of 7 CNNs NIST merged letters (37) affine + elastic 8.21±0.11 Convolutional Neural Network Committees For Handwritten Character Classifications, ICDAR 2011
committee of 7 CNNs NIST letters (case insensitive) affine + elastic 7.58±0.09 Convolutional Neural Network Committees For Handwritten Character Classifications, ICDAR 2011
committee of 7 CNNs NIST letters (case sensitive) affine + elastic 21.41±0.16 Convolutional Neural Network Committees For Handwritten Character Classifications, ICDAR 2011
committee of 7 CNNs NIST all (62) affine + elastic 11.82±0.09 Convolutional Neural Network Committees For Handwritten Character Classifications, ICDAR 2011
35c-MCDNN NIST digits affine + elastic 0.77 Multi-Column Deep Neural Networks For Image Classification, CVPR 2012
35c-MCDNN NIST lowercase letters affine + elastic 7.47 Multi-Column Deep Neural Networks For Image Classification, CVPR 2012
35c-MCDNN NIST uppercase letters affine + elastic 1.83 Multi-Column Deep Neural Networks For Image Classification, CVPR 2012
35c-MCDNN NIST merged letters (37) affine + elastic 7.99 Multi-Column Deep Neural Networks For Image Classification, CVPR 2012
35c-MCDNN NIST letters (case insensitive) affine + elastic 7.37 Multi-Column Deep Neural Networks For Image Classification, CVPR 2012
35c-MCDNN NIST letters (case sensitive) affine + elastic 21.01 Multi-Column Deep Neural Networks For Image Classification, CVPR 2012
35c-MCDNN NIST all (62) affine + elastic 11.63 Multi-Column Deep Neural Networks For Image Classification, CVPR 2012

Handwritten Chinese characters: 3755 classes

NN architecture Dataset Defformations Test Error [%] Info
CNN 11 layers CASIA-HWDB1.1 test set (offline Chinese characters) image size: 48x48 affine + elastic 6.49 Multi-Column Deep Neural Networks For Image Classification, CVPR 2012
NEW! DNN 13 layers CASIA-HWDB1.1 test set (offline Chinese characters) affine + elastic 5.80 June 27th 2013; not published
NEW! 4c-MCDNN CASIA-HWDB1.1 test set (offline Chinese characters) affine + elastic 5.20 May 31st 2013; not published
NEW! 14c-MCDNN CASIA-HWDB1.1 test set (offline Chinese characters) affine + elastic 4.47 August 3rd 2013; not published
NEW! 14c-MCDNN CASIA-HWDB competition test set (offline Chinese characters) - ICDAR 2011 & 2013 affine + elastic 4.22 August 13th 2013; not published; error on best ten predictions - 0.319%
NEW! 8c-MCDNN CASIA-HWDB competition test set (offline Chinese characters) - ICDAR 2011 & 2013 affine + elastic 4.215 August 19th 2013; not published; error on best ten predictions - 0.291%; submitted to competition; tech report
NEW! 4c-MCDNN CASIA-HWDB competition test set (offline Chinese characters) - ICDAR 2011 & 2013 affine + elastic 4.347 August 19th 2013; not published; error on best ten predictions - 0.291%; submitted to competition; tech report

Synthetic images

NN architecture Dataset Distortions IP maps Test Error [%] Info
CNN small NORB no no 7.86±0.55 High-Performance Neural Networks for Visual Object Classification, IJCAI 2011
CNN small NORB 5% translation no 4.71±0.57 High-Performance Neural Networks for Visual Object Classification, IJCAI 2011
CNN small NORB no yes 3.94±0.48 High-Performance Neural Networks for Visual Object Classification, IJCAI 2011
CNN small NORB 5% translation yes 2.53±0.40 High-Performance Neural Networks for Visual Object Classification, IJCAI 2011
CNN committee small NORB 5% translation yes 2.02 Narrowing the Gap to Human Image Classification Performance
DNN cluttered and jittered NORB affine no 3.40±0.23 Multi-Column Deep Neural Networks For Image Classification, CVPR 2012
4c-MCDNN cluttered and jittered NORB affine no 2.70%Multi-Column Deep Neural Networks For Image Classification, CVPR 2012

Natural images

NN architecture Dataset Distortions Test Error [%] Info
8HL CNN CIFAR10 5% translation 19.51±0.18 High-Performance Neural Networks for Visual Object Classification, IJCAI 2011
8HL CNN CIFAR10 affine 16.88±0.33 Narrowing the Gap to Human Image Classification Performance
8 column MCDNN CIFAR10 affine 11.21 Multi-Column Deep Neural Networks For Image Classification, CVPR 2012
10HL CNN CIFAR10 affine 15.88±0.23 -
committee of 10HL CNNs CIFAR10 affine 11.55 -
28-column MCDNN CIFAR10 affine 10.93 -
CNN 8HL CALTECH 101 affine - -
CNN 8HL CALTECH 256 5% translation - -
1 all results were state of the art for their respective benchmark, when published. Some of them still are. I do not work on improving these results as they are either state of the art or very close to state of the art.