Mobile traffic forecasting is an important topic for mobile operators as accurate predictions are essential for en-hancing network efficiency. However, predicting mobile demand is difficult because it requires to distinguish different types of patterns. Recently, deep learning based approaches have been proven effective because of their ability to learn complex non-linear dependencies. Among different architectures, the models built on Recurrent Neural Network (RNN) show impressive performance in sequence modelling and they are widely used in mobile traffic forecasting; however, their rather long training time raises the issue of a high training cost for mobile operators. To better capture the latest trend of mobile demand, mobile op-erators would retrain the forecasting model every week using the recent observations; the training is very expensive and consumes a considerable amount of energy considering that massive cells are managed by mobile operators. For mobile operators, it is always desirable to reduce the training cost without reducing the forecasting performance. To address this challenge, in this paper we propose a pure Fe-layer deep learning predictor allowing the mobile operators to obtain excellent forecasting performance along with very low training cost. The network is composed of three feature extraction layers, containing RNN -inspired blocks learning features that are robust to unexpected local variations; these features are then used to make predictions in the last two layers. Extensive experiments are conduct on a real-world dataset, showing that our model obtains similar prediction performance as the state-of-the-art RNN-based baseline model, while reuuiring only 1 % of its trainina time.

Mobile traffic forecasting plays a key role for optimizing the configuration of network cells. Network operators are very interested in predicting upcoming mobile traffic peaks in an accurate way, in order to improve quality of service via efficient resource allocation. However, forecasting potential peaks is very challenging considering that many peaks occur suddenly for no apparent reason; hence, it is very difficult to determine if a peak will occur in near future based on the temporal dynamics of mobile traffic, potentially leading to inaccurate predictions. To improve the performance of peak prediction, we propose a novel deep learning model called Mixture of Quantiles (MoQ). MoQ employs a mixture of experts model featuring a manager to fuse the predictions of multiple experts. In order to overcome the problem of overly smooth predictions on peaks, the experts are designed to have differentiated forecasting styles from conservative to aggressive. A cooperation mechanism is established through a carefully designed training process, whereby conservative experts are responsible for the forecasting of the off-peak region, and the employed experts are switched to the aggressive ones once the potential increasing trend is detected by manager, which leads to significantly improved peak predictions. Extensive experiments on real-world dataset showcase the effectiveness of the proposed MoQ model, which outperforms all the benchmarks and shows its superior performance in peak forecasting along with excellent interpretability.

Mobile traffic forecasting is crucial for optimizing the network configuration and improving Quality-of-Service (QoS); accurate mobile traffic predictions can help the network operators better configure the network and adapt to the trend of mobile demand. However, improving mobile traffic forecasting is challenging as traffic patterns exhibit strong periodicity while also retaining a certain level of randomness; this makes it difficult to model the sequence correlation. To obtain better mobile traffic predictions, we proposed a deep learning-based predictor called Temporal Dynamics Aware Network (TDANet). TDANet is carefully designed to extract both the global and the local temporal patterns employing recurrent and attention components; a linear module is employed to make the predictions more sensitive to the magnitude of mobile traffic, which makes the model better able to capture fast dynamics. Extensive experiments are conducted on a real-world mobile traffic dataset, and the results show that TDANet outperforms all the baseline models on all evaluation metrics, including both accuracy and complexity.

This work proposes LDVS, a learnable binary local descriptor devised for matching natural images within the MPEG CDVS framework. LDVS descriptors are learned so that they can be sign-quantized and compared using the Hamming distance. The underlying convolutional architecture enjoys a moderate parameters count for operations on mobile devices. Our experiments show that LDVS descriptors perform favorably over comparable learned binary descriptors at patch matching on two different datasets. A complete pair-wise image matching pipeline is then designed around LDVS descriptors, integrating them in the reference CDVS evaluation framework. Experiments show that LDVS descriptors outperform the compressed CDVS SIFT-like descriptors at pair-wise image matching over the challenging CDVS image dataset.

This paper addresses the problem of training a deep neural network for satellite image segmentation so that it can be deployed over images whose statistics differ from those used for training. For example, in postdisaster damage assessment, the tight time constraints make it impractical to train a network from scratch for each image to be segmented. We propose a convolutional encoder-decoder network able to learn visual representations of increasing semantic level as its depth increases, allowing it to generalize over a wider range of satellite images. Then, we propose two additional methods to improve the network performance over each specific image to be segmented. First, we observe that updating the batch normalization layers' statistics over the target image improves the network performance without human intervention. Second, we show that refining a trained network over a few samples of the image boosts the network performance with minimal human intervention. We evaluate our architecture over three data sets of satellite images, showing the state-of-the-art performance in binary segmentation of previously unseen images and competitive performance with respect to more complex techniques in a multiclass segmentation task.

This work addresses the problem of learning compact yet discriminative patch descriptors within a deep learning framework. We observe that features extracted by convolutional layers in the pixel domain are largely complementary to features extracted in a transformed domain. We propose a convolutional network framework for learning binary patch descriptors where pixel domain features are fused with features extracted from the transformed domain. In our framework, while convolutional and transformed features are distinctly extracted, they are fused and provided to a single classifier which thus jointly operates on convolutional and transformed features. We experiment at matching patches from three different dataset, showing that our feature fusion approach outperforms multiple state-of-the-art approaches in terms of accuracy, rate and complexity.

The ever-increasing number of parameters in deep neural networks poses challenges for memory-limited applications. Regularize-and-prune methods aim at meeting these challenges by sparsifying the network weights. In this context we quantify the output sensitivity to the parameters (i.e. their relevance to the network output) and introduce a regularization term that gradually lowers the absolute value of parameters with low sensitivity. Thus, a very large fraction of the parameters approach zero and are eventually set to zero by simple thresholding. Our method surpasses most of the recent techniques both in terms of sparsity and error rates. In some cases, the method reaches twice the sparsity obtained by other techniques at equal error rates.

Convolution is most computationally intensive task of Convolutional Neural Network(CNN). It demands both computational power and memory storage of processing unit. There are different approaches to compute the solution of convolution. In this paper, matrix multiplication based convolution(ConvMM) approach is implemented and accelerated using concurrent resources of Graphics Processing Unit(GPU). CUDA computing language is used to implement this layer. Performance of this GPU-only convolutional layer is compared with its heterogeneous version. Further, flow of this GPU-only convolutional layer is optimized using Unified memory by eliminating overhead caused by extra memory transfers.

We present an Automatic License Plate Recognition system designed around Convolutional Neural Networks (CNNs) and trained over synthetic plate images. We first design CNNs suitable for plate and character detection, sharing a common architecture and training procedure. Then, we generate synthetic images that account for the varying illumination and pose conditions encountered with real plate images and we use exclusively such synthetic images to train our CNNs. Experiments with real vehicle images captured in natural light with commodity imaging systems show precision and recall in excess of 93% despite our networks are trained exclusively on synthetic images.

Fine-grained vehicle classification is a challenging task due to the subtle differences between vehicle classes. Several successful approaches to fine-grained image classification rely on part-based models, where the image is classified according to discriminative object parts. Such approaches require however that parts in the training images be manually annotated, a labor-intensive process. We propose a convolutional architecture realizing a transform network capable of discovering the most discriminative parts of a vehicle at multiple scales. We experimentally show that our architecture outperforms a baseline reference if trained on class labels only, and performs closely to a reference based on a part-model if trained on loose vehicle localization bounding boxes.

Data representation plays an important role in a classifier's accuracy. A given dataset may lead to better results by simply applying a change of basis while keeping the original number of parameters. In this paper, Gabor Filter based image representation has been exploited for object classification. First, Gabor filter based convolution is computed for features extraction, then down-sampling is performed and features are normalized to zero mean and unit variance. This image representation having discriminative visual patterns is used for training of object classifier in Matlab Neural Toolbox. Performance of this proposed image representation is examined on two real world image datasets CIFAR and MNIST and results show that data representation using Gabor can provide good classification without increasing the number of trainable parameters. Finally, this approach is compared to different configurations of Convolutional Neural Network having trainable parameters to verify the validity of proposed image representation.

The choice for image descriptor in a visual navigation system is not straightforward. Descriptors must be distinctive enough to allow for correct localization while still offering low matching complexity and short descriptor size for real-time applications. MPEG Compact Descriptor for Visual Search is a low complexity image descriptor that offers several levels of compromises between descriptor distinctiveness and size. In this work we describe how these trade-offs can be used for efficient loop-detection in a typical indoor environment.

The matching of keypoints present in two images is an uncertain process in which many matches may be incorrect. The statistical properties of the log distance ratio for pairs of in correct matches are distinctly different from the properties of that for correct matches. Based on a statistical model, we propose a goodness-of-fit test in order to establish whether two images contain views of the same object. This technique can be used as a fast geometric consistency check for visual search.

A geographically homogeneous group of citizens shares much common knowledge, characteristics of their culture and history. This knowledge is captured for the use in an item-based recommender system that uses textual information, by introducing bias corpora: newspaper articles that represent the shared knowledge. We present a technique for incorporating and quickly replacing bias corpora in a case study of recommendation of TV contents on our IPTV platform. With this recommender, users watched more items and expressed satisfaction with the service.

The paper provides an overview of the activities that have led to the development of a prototype application of a "talking head" compliant with MPEG-4 specification. The work done so far fits into the context of one of the recent work-items defined by the ISO/IEC JTC1/SC29/WG11, known world-wide as MPEG. This ISO Working Group is now in the process of defining a unifying framework where natural and synthetic audiovisual objects can be combined and rendered as a unique combination of synchronised interactive media. The developed facial animation system implements the following features: animation of predefined/downloaded face models and animation of face models driven either by speech synthesis applications or by MPEG-4 animation parameters. To improve photo-realism, face models can be texture mapped and calibrated according to the countenances of real people.