We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 different classes. On the test data, we achieved top-1 and top-5 error rates of 37.5% and 17.0% which is considerably better than the previous state-of-the-art. The neural network, which has 60 million parameters and 650,000 neurons, consists of five convolutional layers, some of which are followed by max-pooling layers, and three fully-connected layers with a final 1000-way softmax. To make training faster, we used non-saturating neurons and a very efficient GPU implementation of the convolution operation. To reduce overriding in the fully-connected layers we employed a recently-developed regularization method called "dropout" that proved to be very effective. We also entered a variant of this model in the ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%, compared to 26.2% achieved by the second-best entry.

/pdf/imagenet-classification-with-deep-convolutional-neural-1190x9a6b9.pdf

ImageNet Classification with Deep Convolutional Neural Networks

In this work we investigate the effect of the convolutional network depth on its accuracy in the large-scale image recognition setting. Our main contribution is a thorough evaluation of networks of increasing depth using an architecture with very small (3x3) convolution filters, which shows that a significant improvement on the prior-art configurations can be achieved by pushing the depth to 16-19 weight layers. These findings were the basis of our ImageNet Challenge 2014 submission, where our team secured the first and the second places in the localisation and classification tracks respectively. We also show that our representations generalise well to other datasets, where they achieve state-of-the-art results. We have made our two best-performing ConvNet models publicly available to facilitate further research on the use of deep visual representations in computer vision.

Very Deep Convolutional Networks for Large-Scale Image Recognition

The explosion of image data on the Internet has the potential to foster more sophisticated and robust models and algorithms to index, retrieve, organize and interact with images and multimedia data. But exactly how such data can be harnessed and organized remains a critical problem. We introduce here a new database called “ImageNet”, a large-scale ontology of images built upon the backbone of the WordNet structure. ImageNet aims to populate the majority of the 80,000 synsets of WordNet with an average of 500-1000 clean and full resolution images. This will result in tens of millions of annotated images organized by the semantic hierarchy of WordNet. This paper offers a detailed analysis of ImageNet in its current state: 12 subtrees with 5247 synsets and 3.2 million images in total. We show that ImageNet is much larger in scale and diversity and much more accurate than the current image datasets. Constructing such a large-scale database is a challenging task. We describe the data collection scheme with Amazon Mechanical Turk. Lastly, we illustrate the usefulness of ImageNet through three simple applications in object recognition, image classification and automatic object clustering. We hope that the scale, accuracy, diversity and hierarchical structure of ImageNet can offer unparalleled opportunities to researchers in the computer vision community and beyond.

/pdf/imagenet-a-large-scale-hierarchical-image-database-1e8f663s1h.pdf

ImageNet: A large-scale hierarchical image database

We trained a large, deep convolutional neural network to classify the 1.2 million high-resolution images in the ImageNet LSVRC-2010 contest into the 1000 different classes. On the test data, we achieved top-1 and top-5 error rates of 37.5% and 17.0%, respectively, which is considerably better than the previous state-of-the-art. The neural network, which has 60 million parameters and 650,000 neurons, consists of five convolutional layers, some of which are followed by max-pooling layers, and three fully connected layers with a final 1000-way softmax. To make training faster, we used non-saturating neurons and a very efficient GPU implementation of the convolution operation. To reduce overfitting in the fully connected layers we employed a recently developed regularization method called "dropout" that proved to be very effective. We also entered a variant of this model in the ILSVRC-2012 competition and achieved a winning top-5 test error rate of 15.3%, compared to 26.2% achieved by the second-best entry.

/pdf/imagenet-classification-with-deep-convolutional-neural-3kv7rezkq1.pdf

ImageNet classification with deep convolutional neural networks

We introduce a challenging set of 256 object categories containing a total of 30607 images The original Caltech-101 [1] was collected by choosing a set of object categories, downloading examples from Google Images and then manually screening out all images that did not fit the category Caltech-256 is collected in a similar manner with several improvements: a) the number of categories is more than doubled, b) the minimum number of images in any category is increased from 31 to 80, c) artifacts due to image rotation are avoided and d) a new and larger clutter category is introduced for testing background rejection We suggest several testing paradigms to measure classification performance, then benchmark the dataset using two simple metrics as well as a state-of-the-art spatial pyramid matching [2] algorithm Finally we use the clutter category to train an interest detector which rejects uninformative background regions

Caltech-256 Object Category Dataset

Most methods for learning object categories require large amounts of labeled training data. However, obtaining such data can be a difficult and time-consuming endeavor. We have developed a novel, entropy-based ldquoactive learningrdquo approach which makes significant progress towards this problem. The main idea is to sequentially acquire labeled data by presenting an oracle (the user) with unlabeled images that will be particularly informative when labeled. Active learning adaptively prioritizes the order in which the training examples are acquired, which, as shown by our experiments, can significantly reduce the overall number of training examples required to reach near-optimal performance. At first glance this may seem counter-intuitive: how can the algorithm know whether a group of unlabeled images will be informative, when, by definition, there is no label directly associated with any of the images? Our approach is based on choosing an image to label that maximizes the expected amount of information we gain about the set of unlabeled images. The technique is demonstrated in several contexts, including improving the efficiency of Web image-search queries and open-world visual learning by an autonomous agent. Experiments on a large set of 140 visual object categories taken directly from text-based Web image searches show that our technique can provide large improvements (up to 10 x reduction in the number of training examples needed) over baseline techniques.

/pdf/entropy-based-active-learning-for-object-recognition-ktr9i9r6tg.pdf

Entropy-based active learning for object recognition

Probabilistic modelling of text data in the bag-of-words representation has been dominated by directed graphical models such as pLSI, LDA, NMF, and discrete PCA. Recently, state of the art performance on visual object recognition has also been reported using variants of these models. We introduce an alternative undirected graphical model suitable for modelling count data. This "Rate Adapting Poisson" (RAP) model is shown to generate superior dimensionally reduced representations for subsequent retrieval or classification. Models are trained using contrastive divergence while inference of latent topical representations is efficiently achieved through a simple matrix multiplication.

/pdf/the-rate-adapting-poisson-model-for-information-retrieval-xmf41p4eno.pdf

The rate adapting poisson model for information retrieval and object recognition

Learning models for detecting and classifying object categories is a challenging problem in machine vision. While discriminative approaches to learning and classification have, in principle, superior performance, generative approaches provide many useful features, one of which is the ability to naturally establish explicit correspondence between model components and scene features - this, in turn, allows for the handling of missing data and unsupervised learning in clutter. We explore a hybrid generative/discriminative approach using 'Fisher kernels' by Jaakkola and Haussler (1999) which retains most of the desirable properties of generative methods, while increasing the classification performance through a discriminative setting. Furthermore, we demonstrate how this kernel framework can be used to combine different types of features and models into a single classifier. Our experiments, conducted on a number of popular benchmarks, show strong performance improvements over the corresponding generative approach and are competitive with the best results reported in the literature.

/pdf/combining-generative-models-and-fisher-kernels-for-object-yqr5kwq8yl.pdf

Combining generative models and Fisher kernels for object recognition

The topographic projection of retinal ganglion cell (RGC) axons to mouse superior colliculus (SC) or chick optic tectum (OT) is formed in three phases: RGC axons overshoot their termination zone (TZ); they exhibit interstitial branching along the axon that is topographically biased for the correct location of their future TZ; and branches arborize preferentially at the TZ and the initial exuberant projection refines through axon and branch elimination to generate a precise retinotopic map. We present a computational model of map development that demonstrates that the countergradients of EphAs and ephrinAs in retina and the OT/SC and bidirectional repellent signaling between RGC axons and OT/SC cells are sufficient to direct an initial topographic bias in RGC axon branching. Our model also suggests that a proposed repellent action of EphAs/ephrinAs present on RGC branches and arbors added to that of EphAs/ephrinAs expressed by OT/SC cells is required to progressively restrict branching and arborization to topographically correct locations and eliminate axon overshoot. Simulations show that this molecular framework alone can develop considerable topographic order and refinement, including axon elimination, a feature not programmed into the model. Generating a refined map with a condensed TZ as in vivo requires an additional parameter that enhances branch formation along an RGC axon near sites that it has a higher branch density, and resembles an assumed role for patterned neural activity. The same computational model generates the phenotypes reported in ephrinA deficient mice and Isl2-EphA3 knockin mice. This modeling suggests that gradients of counter-repellents can establish a substantial degree of topographic order in the OT/SC, and that repellents present on RGC axon branches and arbors make a substantial contribution to map refinement. However, competitive interactions between RGC axons that enhance the probability of continued local branching are required to generate precise retinotopy.

/pdf/computational-modeling-of-retinotopic-map-development-to-4rwir4ydgs.pdf

Computational Modeling of Retinotopic Map Development to Define Contributions of EphA-EphrinA Gradients, Axon-Axon Interactions, and Patterned Activity

Alex D. Holub

Papers

Caltech-256 Object Category Dataset

Entropy-based active learning for object recognition

The rate adapting poisson model for information retrieval and object recognition

Combining generative models and Fisher kernels for object recognition

Computational Modeling of Retinotopic Map Development to Define Contributions of EphA-EphrinA Gradients, Axon-Axon Interactions, and Patterned Activity