Convolutional neural networks applied to house numbers digit classification

By Pierre Sermanet, Soumith Chintala, Yann LeCun. 2012

566 citations recorded by the source corpus.

We classify digits of real-world house numbers using convolutional neural networks (ConvNets). Con-vNets are hierarchical feature learning neural networks whose structure is biologically inspired. Unlike many popular vision approaches that are hand-designed, ConvNets can automatically learn a unique set of features optimized for a given task. We augmented the traditional ConvNet architecture by learning multi-stage features and by using Lp pooling and establish a new state-of-the-art of 95.10% accuracy on the SVHN dataset (48% error improvement). Furthermore, we analyze the benefits of different pooling methods and multi-stage features in ConvNets. The source code and a tutorial are available at eblearn.sf.net.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Pylearn2: a machine learning research library (2013, methodology): “Additionally, there are many kinds of preprocessing, such as PCA [38], ZCA [4], various kinds of local contrast normalization [46], as well as helper functions to set up the entire preprocessing p ipeline from some well-known successful and documented systems [8, 45].”
Spectral Networks and Locally Connected Networks on Graphs (2013, background): “Speech [10], images [12, 18, 20] or video [21, 16] are prominent examples that fall into this category.”
Holistically-Nested Edge Detection (2015, background): “Skip-layer network learning: Examples of this form of network include [26, 14, 2, 33, 10].”
Densely Connected Convolutional Networks (2016, methodology): “Following common practice [7, 13, 20, 22, 29] we use all the training data without any data augmentation, and a validation set with 6,000 images is split from the training set.”
A Unified Gradient Regularization Family for Adversarial Examples (2015, methodology): “As [10] have indicated that the extreme case of hyper parameter p (p = 1 or p =∞) might not be the optimal setting, we would like to introduce standard gradient regularization where p = 2.”
Designing Neural Network Architectures using Reinforcement Learning (2016, methodology): “These layers generate N learnable linear transformations of the input data, which is similar in spirit to preprocessing of input data from RGB to a different color spaces such as YUV, as found in prior work (Sermanet et al., 2012; 2013).”
Compression of fully-connected layer in neural network by Kronecker product (2015, methodology): “To build a validation set, we randomly select 400 images per class from training set and 200 images per class from extra training set as Sermanet et al. (2012), Goodfellow et al.”
Stochastic Pooling for Regularization of Deep Convolutional Neural Networks (2013, methodology): “Net + 2-layer Classifer + padding [12] – 4.”
Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks (2013, methodology): “, 2001), local contrast normalization (Sermanet et al., 2012), etc.”
Snapshot Ensembles: Train 1, get M for free (2017, methodology): “Following common practice (Sermanet et al., 2012; Goodfellow et al., 2013; Huang et al., 2016a), we withhold 6,000 training images for validation, and train on the remaining images without data augmentation.”