Anastasios Tefas

Deep Learning Analytics

The recent breakthroughs in Deep Learning have provided powerful data analytics tools for a wide range of domains ranging from advertising and analyzing users’ behavior to load and financial forecasting. Depending on the nature of the available data and the task at hand Deep Learning Analytics techniques can be divided into two broad categories: (a) unsupervised learning techniques and (b) supervised learning techniques. In this chapter we provide an extensive overview over both categories. Unsupervised learning methods, such as Autoencoders, are able to discover and extract the information from the data without using any ground truth information and/or supervision from domain experts. Thus, unsupervised techniques can be especially useful for data exploration tasks, especially when combined with advanced visualization techniques. On the other hand, supervised learning techniques are used when ground truth information is available and we want to build classification and/or forecasting models. Several deep learning models are examined ranging from simple Multilayer Perceptrons (MLPs) to Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). However training deep learning models is not always a straightforward task requiring both a solid theoretical background as well as intuition and experience. To this end, we also present recent techniques that allow for efficiently training deep learning models, such as batch normalization, residual connections, advanced optimization techniques and activation functions, as well as a number of useful practical suggestions. Finally, we present an overview of the available open source deep learning frameworks that can be used to implement deep learning analytics techniques and accelerate the training process using Graphics Processing Units (GPUs).

Recurrent Attention for Deep Neural Object Detection

Recent advances in deep learning have achieved state-of-the-art results for object detection by replacing the traditional detection methodologies with deep convolutional neural network architectures. A contemporary technique that is shown to further improve the performance of these models on tasks ranging from optical character recognition and neural machine translation to object detection is based on incorporating an attention mechanism within the models. The idea behind the attention mechanism and its variations was to improve the information quality extracted for any confronted task by focusing on the most relevant parts of the input. In this paper we propose two novel deep neural architectures for object recognition that incorporate the idea of the attention mechanism in the well-known faster-RCNN object detector. The objective is to develop attention mechanisms that can be used for small objects detection as they appear when using Drones for covering sport events like bicycle races, football matches and rowing races. The proposed approaches include a class agnostic method that applies the same predetermined context for every class, and a class specific method which learns to include context that maximizes the classs precision individually for each class. The proposed methods are evaluated in the VOC2007 dataset, improving the performance of the baseline faster-RCNN architecture.

Deep Sparse Autoencoders for Football Match and Bet Prediction

In recent years, betting in Association Football has become increasingly popular. A lot of research has been done on the subject using statistical and other approaches to create predictive models. Correctly forecasting a match outcome is a difficult and complex problem to tackle, due to the nature of the sport. There is a plethora of data to work with, such as odds by bookmakers, team stats, players performance, etc. With the recent resurgence of Artificial Neural Networks, we try to apply their predictive ability to this problem using Sparse Deep Autoencoders, taking advantage of the odds given by the bookmakers. Our main objective is to get the best possible predictions of the probabilities of each one of the three outcomes of a football match. From that, we can either bet on the highest probability or use the probabilities to calculate the expected value of each bet, i.e. the expected monetary return of it, aiming for the highest accuracy possible while simultaneously making a profit at the end of each season, all the while following a consistent betting strategy.

Learning Deep Representations with Probabilistic Knowledge Transfer

Knowledge Transfer (KT) techniques tackle the problem of transferring the knowledge from a large and complex neural network into a smaller and faster one. However, existing KT methods are tailored towards classification tasks and they cannot be used efficiently for other representation learning tasks. In this paper we propose a novel probabilistic knowledge transfer method that works by matching the probability distribution of the data in the feature space instead of their actual representation. Apart from outperforming existing KT techniques, the proposed method allows for overcoming several of their limitations providing new insight into KT as well as novel KT applications, ranging from KT from handcrafted feature extractors to cross-modal KT from the textual modality into the representation extracted from the visual modality of the data.

Computational UAV Cinematography for Intelligent A/V Shooting Based on Semantic Visual Analysis

As audiovisual coverage of sports events using Unmanned Aerial Vehicles (UAVs) is becoming increasingly popular, intelligent audiovisual (A/V) shooting tools are needed to assist the cameramen and directors. Several challenges also arise by employing autonomous UAVs, including the accurate identification of the 2D region of cinematographic attention (RoCA) depicting rapidly moving target ensembles (e.g., athletes) and the automatic control of the UAVs so as to take informative and aesthetically pleasing A/V shots, by performing automatic or semiautomatic visual content analysis with no or minimal human intervention. A novel method implementing computational UAV cinematography for assisting sports coverage, based on semantic, human-centered visual analysis is proposed in this work. Athlete detection and tracking, as well as spatial athlete distribution on the image plane are the semantic features extracted from an aerial video feed captured by a UAV and exploited for the extraction of the RoCA, based solely on present and past athlete detections and their regions of interest (ROIs). A PID controller that visually controls a real or virtual camera in order to track the sports RoCA and produce aesthetically pleasing shots, without using 3D location-related information, is subsequently employed. The proposed method is evaluated on actual UAV A/V footage from soccer matches and promising results are obtained.

Learning deep spatiotemporal features for video captioning

Abstract In this paper, we propose a novel automatic video captioning system which translates videos to sentences, utilizing a deep neural network that is composed of three building parts of convolutional and recurrent structure. That is, the first subnetwork operates as feature extractor of single frames. The second subnetwork is a three-stream network, capable of capturing spatial semantic information in the first stream, temporal semantic information in the second stream, and global video concept information in the third stream. The third subnetwork generates relevant textual captions using as input the spatiotemporal features of the second subnetwork. The experimental validation indicates the effectiveness of the proposed model, achieving superior performance over competitive methods.

Long-term temporal averaging for stochastic optimization of deep neural networks

Deep learning models are capable of successfully tackling several difficult tasks. However, training deep neural models is not always a straightforward task due to several well-known issues, such as the problems of vanishing and exploding gradients. Furthermore, the stochastic nature of most of the used optimization techniques inevitably leads to instabilities during the training process, even when state-of-the-art stochastic optimization techniques are used. In this work, we propose an advanced temporal averaging technique that is capable of stabilizing the convergence of stochastic optimization for neural network training. Six different datasets and evaluation setups are used to extensively evaluate the proposed method and demonstrate the performance benefits. The more stable convergence of the algorithm also reduces the risk of stopping the training process when a bad descent step was taken and the learning rate was not appropriately set.

Big Media Data Analysis

Abstract In this editorial a short introduction to the special issue on Big Media Data Analysis is given. The scope of this Editorial is to briefly present methodologies, tasks and applications of big media data analysis and to introduce the papers of the special issue. The special issue includes six papers that span various media analysis application areas like generic image description, medical image and video analysis, distance calculation acceleration and data collection.

2014 IEEE International Conference on Image Processing (ICIP 2014)

In this paper, we propose a framework for clustering shots from stereoscopic videos into clusters that correspond to semantic concepts exploiting visual and disparity information. Various color, disparity and texture descriptors are applied to shot key frames for obtaining low-level representations. Self Organizing Maps are subsequently employed upon various combinations of these representations in order to determine a lattice of representative semantic concepts. Experimental results on performances and football stereoscopic videos show that the use of disparity information leads to better clustering compared to using visual information only.

2014 IEEE Symposium on Computational Intelligence in Big Data (CIBD 2014)

With the proliferation of the World Wide Web, graph structures have arisen on social network/media sites. Such graphs usually number several million nodes, i.e., they can be characterized as Big Data. Graph clustering is an important analysis tool for other graph related tasks, such as compression, community discovery and recommendation systems, to name a few. We propose a novel extension to a graph clustering algorithm, that attempts to cluster a graph, through the optimization of selected terms of the graph weight/adjacency matrix Discrete Cosine Transform.