Computer Vision with Fast & Deep / Recurrent Neural Nets: Best Results. By Juergen Schmidhuber. Includes adapted HAL 9000 image from Kubrick's movie based on CLarke's Novel: 2001. Also includes inmages from the German Traffic Sign Recognition Benchmark.

Deep Learning: Fast Deep / Recurrent Neural Networks Win Many Computer Vision Contests for Object Detection, Image Segmentation, Pattern Classification
Jürgen Schmidhuber, 2009-2013 (compare 2012 interview on KurzweilAI)

Computer vision and pattern recognition is becoming essential for thousands of practical applications. For example, the future of search engines lies in image and video recognition as opposed to traditional text search. Autonomous robots such as driverless cars depend on it, too. It even has lifesaving impact through medical applications such as cancer detection.

Since 2009, our Deep Learning team has won 9 (nine) first prizes in important and highly competitive international pattern recognition contests (with test set known only to the organisers), far more than any other team. Our neural nets also were the first really Deep Learners to win such contests (e.g., on image classification, object detection, image segmentation), and the first machine learning methods to reach superhuman performance in such a contest. Here the list of won competitions (details in the rightmost column):

9. MICCAI 2013 Grand Challenge on Mitosis Detection
8. ICPR 2012 Contest on Mitosis Detection in Breast Cancer Histological Images
7. ISBI 2012 Brain Image Segmentation Challenge (with superhuman pixel error rate)
6. IJCNN 2011 Traffic Sign Recognition Competition (only our method achieved superhuman results)
5. ICDAR 2011 offline Chinese Handwriting Competition
4. Online German Traffic Sign Recognition Contest
3. ICDAR 2009 Arabic Connected Handwriting Competition
2. ICDAR 2009 Handwritten Farsi/Arabic Character Recognition Competition
1. ICDAR 2009 French Connected Handwriting Competition. Compare the overview page on handwriting recognition.

Our deep learning methods also set records in important Machine Learning (ML) benchmarks (details in the rightmost column):

A. The NORB Object Recognition Benchmark
B. The CIFAR Image Classification Benchmark
C. The MNIST Handwritten Digits Benchmark (perhaps the most famous benchmark; we achieved the 1st human-competitive result in 2011)
D. Chinese characters from the ICDAR 2013 competition (3755 classes)
E. The Weizmann & KTH Human Action Recognition Benchmarks

Remarkably, none of 1-9 & A-D above required the traditional sophisticated computer vision techniques developed over the past six decades or so. Instead, our biologically rather plausible systems are inspired by human brains, and learn to recognize objects from numerous training examples. We use deep, artificial, supervised, feedforward or recurrent (deep by nature) neural networks with many non-linear processing stages.

We started work on "deep learning" over two decades ago. Back then, Sepp Hochreiter (now professor) was an undergrad student working on Schmidhuber's neural net project. His 1991 thesis (PDF) is a deep learning milestone: it formally showed that deep networks like the above are hard to train because they suffer from the now famous problem of vanishing or exploding gradients. Since then we have developed various techniques to overcome this obstacle (e.g., see here). In 1991 we used a stack of recurrent neural networks (RNN) pre-trained in unsupervised fashion to compactly encode input sequences, where lower layers of the hierarchy learn to extract compact sequence representations fed to higher layers. This can greatly accelerate subsequent supervised learning. See J. Schmidhuber, Learning complex, extended sequences using the principle of history compression, Neural Computation, 4(2):234-242, 1992, PDF. (Based on TR FKI-148-91, 1991.) See Schmidhuber's Habilitation thesis (TUM, 1993) for an experiment with 1200 nonlinear virtual layers. The 1991 system also was the first Neural Hierarchical Temporal Memory.

Today we use graphics cards or GPUs (mini-supercomputers for video games, see picture in 2nd column) to speed up learning by a factor of up to 50. Our committees of networks improve the results even further.

Note that the successes 1-9 & A-D above did NOT require any unsupervised pre-training, which is a bit depressing, as we have developed unsupervised learning algorithms for 20 years. However, our new systems' feature detectors (FDs) do resemble those found by our old unsupervised methods such as Predictability Minimization (1992; more: 1996) and Low-Complexity Coding and Decoding (LOCOCODE, 1999).

Our new systems' feature detectors resemble 
  those found by our old unsupervised methods such as
Predictability Minimization (1992). Jürgen Schmidhuber FD due to semi- linear PM (1992), made in 1996

Reference [14] also uses fast deep nets, this time to achieve superior hand gesture recognition. And reference [16] uses them to achieve superior steel defect detection, three times better than support vector machines (SVM) trained on commonly used feature descriptors.

Not all of our pattern recognizers use neural nets though. For E above, novel supervised and unsupervised kernel-based methods were employed. Reference [5] uses credal classifiers to classify textures. And reference [2] uses a fast voting scheme to answer image-based queries, successfully tested on the ZuBud database.

For information on how we have built on the work of earlier pioneers since the 1960s, please visit www.deeplearning.me.

Copyright notice (2011): Fibonacci web design by Jürgen Schmidhuber, who will be delighted if you use this web page for educational and non-commercial purposes, including articles for Wikipedia and similar sites, provided you mention the source and provide a link. The Deep Learning site derives from the present page.

Last update 2014

NORB Dataset: Best Results as of 2011, by Fast Deep Nets on GPUs. Segmentation of neuronal structures in EM stacks: Best Results as of 2012. (Juergen Schmidhuber)

Our simple training algorithms for deep, wide, artificial neural network architectures similar to those of biological brains now win many competitions and yield best known results on many famous benchmarks for visual pattern recognition. Shown here are example images from NORB and EM stacks (left), CIFAR-10, Weizmann, KTH (below), and the Traffic Sign Competition (above / below).

We are currently experiencing a second Neural Network ReNNaissance (title of JS' IJCNN 2011 keynote) - the first one happened in the 1980s and early 90s. In many applications, our deep NNs are now outperforming all other methods including the theoretically less general and less powerful support vector machines (which for a long time had the upper hand, at least in practice). Check out the, in hindsight, not too optimistic predictions of our RNNaissance workshop at NIPS 2003, and compare the RNN book preface.

Computer Vision Team (ex-)members in Schmidhuber's lab(s): Dan Ciresan, Ueli Meier, Jonathan Masci, Somayeh Danafar, Alex Graves, Davide Migliore. For medical imaging, we also work with Alessandro Giusti in the group of Luca Maria Gambardella.

Our work builds on earlier work by great neural network pioneers including Bryson, Kelley, Dreyfus, Werbos, Fukushima, Amari, LeCun, Hinton, Williams, Rumelhart, Poggio, von der Malsburg, Kohonen, and others (more).

GPUs used for the work described in: High-Performance Neural Networks for Visual Object Classification, arXiv:1102.0183v1 [cs.AI]. (Jürgen Schmidhuber) SELECTED PUBLICATIONS

[24] J. Schmidhuber. Deep Learning in Neural Networks: An Overview. Neural Networks, Volume 61, January 2015, Pages 85-117, published online in 2014 (DOI: 10.1016/ j.neunet.2014.09.003). Draft of invited survey (88 pages, 888 references): Preprint IDSIA-03-14 / arXiv:1404.7828 [cs.NE]; version v4 (PDF, 8 Oct 2014); LATEX source; complete public BIBTEX file (888 kB). (Older PDF versions: v1 of 30 April; v1.5 of 15 May; v2 of 28 May; v3 of 2 July.) HTML overview page.

[23] D. Ciresan, J. Schmidhuber. Multi-Column Deep Neural Networks for Offline Handwritten Chinese Character Classification. Preprint arXiv:1309.0261, 1 Sep 2013.

[22] J. Masci, A. Giusti, D. Ciresan, G. Fricout, J. Schmidhuber. A Fast Learning Algorithm for Image Segmentation with Max-Pooling Convolutional Networks. ICIP 2013. Preprint arXiv:1302.1690. On object detection in large images, now scanned by our deep networks 1500 times faster than with previous methods.

[21] A. Giusti, D. Ciresan, J. Masci, L.M. Gambardella, J. Schmidhuber. Fast Image Scanning with Deep Max-Pooling Convolutional Neural Networks. ICIP 2013. Preprint arXiv:1302.1700

[20] D. Ciresan, A. Giusti, L. M. Gambardella, J. Schmidhuber. Mitosis Detection in Breast Cancer Histology Images using Deep Neural Networks. MICCAI 2013. PDF.

[19] D. Ciresan, U. Meier, J. Schmidhuber. Transfer Learning for Latin and Chinese Characters with Deep Neural Networks. Proc. IJCNN 2012, p 1301-1306, 2012. PDF. Pretrain on one data set, profit on another.

[18] D. C. Ciresan, U. Meier, J. Masci, J. Schmidhuber. Multi-Column Deep Neural Network for Traffic Sign Classification. Neural Networks 32, p 333-338, 2012. PDF of preprint. (First superhuman visual pattern recognition.)

[17] D. C. Ciresan, U. Meier, J. Schmidhuber. Multi-column Deep Neural Networks for Image Classification. IEEE Conf. on Computer Vision and Pattern Recognition CVPR 2012, p 3642-3649, 2012. PDF. Longer preprint arXiv:1202.2745v1 [cs.CV].

[16] J. Masci, U. Meier, D. Ciresan, G. Fricout, J. Schmidhuber. Steel Defect Classification with Max-Pooling Convolutional Neural Networks. Proc. IJCNN 2012. PDF.

[15] D. Ciresan, A. Giusti, L. Gambardella, J. Schmidhuber. Deep Neural Networks Segment Neuronal Membranes in Electron Microscopy Images. In Advances in Neural Information Processing Systems (NIPS 2012), Lake Tahoe, 2012. PDF. (See also ISBI EM Competition Abstracts.)

[14] J. Nagi, F. Ducatelle, G. A. Di Caro, D. Ciresan, U. Meier, A. Giusti, F. Nagi, J. Schmidhuber, L. M. Gambardella. Max-Pooling Convolutional Neural Networks for Vision-based Hand Gesture Recognition. Proc. 3rd IEEE Intl. Conf. on Signal & Image Processing and Applications (ICSIPA), Kuala Lumpur, 2011. PDF.

[13] J. Schmidhuber, D. Ciresan, U. Meier, J. Masci, A. Graves. On Fast Deep Nets for AGI Vision. In Proc. Fourth Conference on Artificial General Intelligence (AGI-11), Google, Mountain View, California, 2011. PDF. Video.

[12] D. C. Ciresan, U. Meier, L. M. Gambardella, J. Schmidhuber. Convolutional Neural Network Committees For Handwritten Character Classification. 11th International Conference on Document Analysis and Recognition (ICDAR 2011), Beijing, China, 2011. PDF.

[11] U. Meier, D. C. Ciresan, L. M. Gambardella, J. Schmidhuber. Better Digit Recognition with a Committee of Simple Neural Nets. 11th International Conference on Document Analysis and Recognition (ICDAR 2011), Beijing, China, 2011. PDF.

[10] D. C. Ciresan, U. Meier, J. Masci, J. Schmidhuber. A Committee of Neural Networks for Traffic Sign Classification. International Joint Conference on Neural Networks (IJCNN-2011, San Francisco), 2011. PDF.

[9] D. C. Ciresan, U. Meier, L. M. Gambardella, J. Schmidhuber. Handwritten Digit Recognition with a Committee of Deep Neural Nets on GPUs. ArXiv Preprint arXiv:1103.4487v1 [cs.LG], 23 Mar 2011.

[8] D. C. Ciresan, U. Meier, J. Masci, L. M. Gambardella, J. Schmidhuber. Flexible, High Performance Convolutional Neural Networks for Image Classification. International Joint Conference on Artificial Intelligence (IJCAI-2011, Barcelona), 2011. PDF. ArXiv preprint, 1 Feb 2011. Describes our special breed of max-pooling convolutional networks (MPCNN), now widely tested/used by research labs (e.g., Univ. Toronto/Google/Stanford) and companies (e.g., Apple) all over the world.

[7] S. Danafar, A. Giusti, J. Schmidhuber. New State-of-the-Art Recognizers of Human Actions. EURASIP Journal on Advances in Signal Processing, doi:10.1155/2010/202768, 2010. HTML.

[6] D. C. Ciresan, U. Meier, L. M. Gambardella, J. Schmidhuber. Deep Big Simple Neural Nets For Handwritten Digit Recognition. Neural Computation 22(12): 3207-3220, 2010. ArXiv Preprint arXiv:1003.0358v1 [cs.NE], 1 March 2010.

[5] G. Corani, A. Giusti, D. Migliore, J. Schmidhuber. Robust Texture Recognition Using Credal Classifiers. Proc. BMVC, p 78.1-78.10. BMVA Press, 2010. doi:10.5244/C.24.78. HTML.

[4] A. Graves, J. Schmidhuber. Offline Handwriting Recognition with Multidimensional Recurrent Neural Networks. Advances in Neural Information Processing Systems 22, NIPS'22, p 545-552, Vancouver, MIT Press, 2009. PDF.

[3] A. Graves, S. Fernandez, J. Schmidhuber. Multi-Dimensional Recurrent Neural Networks. Intl. Conf. on Artificial Neural Networks ICANN'07, 2007. Preprint: arxiv:0705.2011. PDF.

[3a] A. Graves, S. Fernandez, F. Gomez, J. Schmidhuber. Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks. ICML 06, Pittsburgh, 2006. PDF.

[2] M.v.d. Giessen and J. Schmidhuber. Fast color-based object recognition independent of position and orientation. In W. Duch et al. (Eds.): Proc. Intl. Conf. on Artificial Neural Networks ICANN'05, LNCS 3696, pp. 469-474, Springer-Verlag Berlin Heidelberg, 2005. PDF.

Link to the first artificial fovea sequentially steered by a learning neural controller (1990). Jürgen Schmidhuber Ongoing work on active perception. While the methods above tend to work fine in many applications, they are passive learners - they do not learn to actively search for the most informative image parts. Humans, however, use sequential gaze shifts for pattern recognition. This can be much more efficient than the fully parallel one-shot approach. That's why we want to combine the algorithms above with variants of our old method of 1990 - back then we built what to our knowledge was the first artificial fovea sequentially steered by a learning neural controller. Without a teacher, it used a variant of reinforcement learning to create saccades and find targets in a visual scene (and to track moving targets), although computers were a million times slower back then:

[1] J. Schmidhuber, R. Huber. Learning to generate artificial fovea trajectories for target detection. International Journal of Neural Systems, 2(1 & 2):135-141, 1991 (figures omitted). PDF. HTML. HTML overview with figures.

More on active learning without a teacher in the overview pages on the Formal Theory of Creativity and Curiosity.

More Deep Learning Web Sites:

Deep Learning since 1991 (overview site derived from the present page)

Sept/Oct 2013: G+ posts on Deep Learning

Deep NN win MICCAI 2013 Grand Challenge and 2012 ICPR Contest on Mitosis Detection (first Deep Learner to win a contest on object detection in large images)

Deep NN win 2012 Brain Image Segmentation Contest (first image segmentation competition won by a feedforward Deep Learner)

2012: 8th international pattern recognition contest won since 2009

2011: First superhuman visual pattern recognition (twice better than humans, three times better than the closest artificial competitor, six times better than the best non-neural method)

2009: First official international pattern recognition contests won by Deep Learning (connected handwriting through LSTM RNN: simultaneous segmentation and recognition)

1997: First purely supervised Deep Learner (LSTM RNN)

JS' first Deep Learner of 1991 + Deep Learning Timeline 1962-2013 (also summarises the origins of backpropagation, still the central algorithm of Deep Learning)

1991: Fundamental Deep Learning Problem discovered and analysed and partially solved

CIFAR Dataset: Best Results as of  2011, by Fast Deep Nets on GPUs. (Jürgen Schmidhuber)


Links to the original datasets of competitions and benchmarks, plus more information on the world records set by our team:

16. 22 Sept 2013: our deep and wide MCMPCNN [8,17] won the MICCAI 2013 Grand Challenge on Mitosis Detection (important for cancer prognosis etc). This was made possible through the efforts of Dan and Alessandro [20]. Don't confuse this with the earlier ICPR 2012 Contest below! Comment: When we started our work on deep learning over two decades ago, limited computing power forced us to focus on tiny toy applications to illustrate the benefits of our methods. How things have changed! It is gratifying to observe that today our techniques may actually help to improve healthcare and save lives.

15. As of 1 Sep 2013, our Deep Learning Neural Networks are the best artificial offline recognisers of Chinese characters from the ICDAR 2013 competition (3755 classes), approaching human performance [23]. This is relevant for smartphone producers who want to build phones that can translate photos of foreign texts and signs. As always in such competitions, GPU-based pure supervised gradient descent (40-year-old backprop) was applied to deep and wide multi-column networks with interleaving max-pooling layers and convolutional layers (multi-column MPCNN) [8,17]. Many leading IT companies and research labs are now using this technique, too.

14. ICPR 2012 Contest on Mitosis Detection in Breast Cancer Histological Images (MITOS Aperio images). There were 129 registered companies / institutes / universities from 40 countries, and 14 results. Our team (with Alessandro & Dan) clearly won the contest (over 20% fewer errors than the second best team). See ref [20], as well as the later MICCAI 2013 Grand Challenge above.

13. ISBI 2012 Segmentation of neuronal structures in EM stacks challenge. See the TrakEM2 data sets of INI. Our team won the contest on all three evaluation metrics by a large margin, with superhuman performance in terms of pixel error (March 2012) [15]. (Ranks 2-6 for researchers at ETHZ, MIT, CMU, Harvard.)
This is relevant for the recent huge brain projects in Europe and the US, which try to build 3D models of real brains.

12. IJCNN 2011 on-site Traffic Sign Recognition Competition (1st rank, 2 August 2011, 0.56% error rate, the only method better than humans, who achieved 1.16% on average; 3rd place for 1.69%) [10,18]. The first method ever to achieve superhuman visual pattern recognition on an important benchmark (with deadline and test set known only to the organisers). This is obviously relevant for self-driving cars.

11. INI @ Univ. Bochum's online German Traffic Sign Recognition Benchmark, won through late night efforts of Dan & Ueli & Jonathan (1st & 2nd rank; 1.02% error rate, January 2011) [10].

10. NORB object recognition dataset for stereo images, NY University, 2004. Our team set the new record on the standard set (2.53% error rate) in January 2011 [8], and achieved 2.7% on the full set [17] (best previous result by others: 5%).

9. The CIFAR-10 dataset of Univ. Toronto, 2009. Our team set the new record (19.51% error rate) on these rather challenging data in January 2011 [8], and improved this to 11.2% [17].

IJCNN 2011 on-site Traffic Sign Recognition Competition (1st rank, 2 August 2011, 0.56% error rate, the only method better than humans, who achieved 1.16% on average; 3rd place for 1.69%) (Juergen Schmidhuber)

8. The MNIST dataset of NY University, 1998. Our team set the new record (0.35% error rate) in 2010 [6], tied it again in January 2011 [8], broke it again in March 2011 (0.31%) [9], and again (0.27%, ICDAR 2011) [12], and finally achieved the first human-competitive result: 0.23% [17] (mean of many runs; many individual runs yield better results, of course, down to 0.17% [12]).

7. The Chinese Handwriting Recognition Competition at ICDAR 2011 (offline). Our team won 1st and 2nd rank (CR(1): 92.18% correct; CR(10): 99.29% correct) in June 2011.

Three Connected Handwriting Recognition Competitions at ICDAR 2009 were won by our multi-dimensional LSTM recurrent neural networks [3,3a,4] through the efforts of Alex. This was the first RNN system ever to win an official international pattern recognition competition. To our knowledge, this also was the first Deep Learning system ever (recurrent or not) to win such a contest:

6. ICDAR 2009 Arabic Connected Handwriting Competition of Univ. Braunschweig

5. ICDAR 2009 Handwritten Farsi/Arabic Character Recognition Competition

4. ICDAR 2009 French Connected Handwriting Competition (PDF) based on data from the RIMES campaign

Note that 4-8 are treated in more detail in the page on handwriting recognition.

KTH and Weizmann Datasets: Best Results as of 2010, by Kernel-Based Recognizers of Human Actions. (Jürgen Schmidhuber)

3. The Weizmann Human Action Dataset of Weizmann Institute of Science, and the KTH Human Action Dataset of KTH Royal Insitute of Technology. New records set in 2010 [7], thanks to Somayeh's efforts.

2. The Outex Texture Database, Univ. Oulu, 2002 [5].

1. The ZuBuD database of pictures of buildings in Zürich, ETHZ, 2003 [2].

Here a 12 min Google Tech Talk video on fast deep / recurrent nets (only slides and voice) at AGI 2011, summarizing results as of August 2011:

Google Tech Talk video (13:05) on fast deep / recurrent neural networks for computer vision presented by Juergen Schmidhuber at AGI 2011 at Google HQ, Mountain View, CA.

People keep asking: What is the secret of your successes? There are two secrets:

(i) For competitions involving sequential data such as video and speech we use deep (stacks of) multi-dimensional [3] Long Short-Term Memory (LSTM) recurrent networks (1997) trained by Connectionist Temporal Classification (CTC, 2006) [3a]. This is what since 2009 has set records in recognising connected handwriting and speech.

(ii) For other competitions we use multi-column committees [10] of GPU-based max-pooling CNN (2011) [8], where we apply (in the style of LeCun et al 1989, Ranzato et al 2007) efficient backpropagation (Linnainmaa 1970, Werbos 1981) to deep Neocognitron-like weight-sharing convolutional architectures (Fukushima 1979) with max-pooling layers (Weng 1992, Riesenhuber & Poggio 1999). Over two decades, LeCun's lab has invented many improvements of such CNN. Our GPU-MPCNN achieved the first superhuman image recognition results (2011) [18], and were the first Deep Learners to win contests in object detection (2012) and image segmentation (2012), which require fast, non-redundant MPCNN image scans [21,22].

Our algorithms not only were the first deep learning methods to win official international competitions (since 2009) and to become human-competitive, they also have numerous immediate industrial and medical applications. Apple & Google and many others adopted our techniques. Are you an industrial company that wants to solve interesting pattern recognition problems? Don't hesitate to contact JS. We already developed:

1. State-of-the-art handwriting recognition for a software services company.
2. State-of-the-art steel defect detection for the world's largest steel maker.
3. State-of-the-art low-cost pattern recognition for a leading automotive supplier.
4. Low-power variants of our methods for apps running on cell phone chips.