Online Search Campaign Management using Latent Topic Models
In this project, a digital campaign is represented by its main and distinct themes, which are identified by Probabilistic Inference. This is contrary to a manual evaluation at keyword phrase-level. If each theme is represented by a different color, a campaign can be thought of as a portrait composed out of various colors in different proportions. Since each color represents a focused theme, an advertiser can shape his/her campaign with her color palette on the fly. For example, if more sales took place in the months dominated by red color, the weight of red can be increased by the computed generative process. Following the well known motto "a picture is worth a thousand words", we can, with the use of this color coding, arrive at the following conclusions: (i) more users have visited the website in those months dominated by blue color, (ii) users stayed longer in the site on those days dominated by green color.Principle Investigator: Ahmet Bulut
Smart Living Assistance Project
In this project we want to develop a SLAS that recognizes a person’s daily routines, and lets the person interact with the devices in its environment naturally. To date, computers have limited means of communicating with humans. Most common methods involve using a tethered device (keyboard, mouse, etc.) that limits the user’s freedom of motion. With the advent of more powerful computers equipped with video and depth cameras, vision based interfaces are becoming more feasible. Enabling the computer to “see” its user allows for richer and more varied paradigms of human-machine interaction.In our smart living assistance system the user will be able to interact with the devices in the environment to perform daily tasks via gestures by moving his/her hands and fingers. The user will not need to wear gloves or any sensors but will be continously monitored with our depth cameras to recognize predefined patterns i.e., gestures. These gesture will let the user to control the smart TV, the air-conditioner, open up the mail application in his computer and browse through his emails, and more. Gesture recognition has wide-ranging applications such as navigating in virtual environments, controlling devices in a smart environment, interacting with computer games via player’s gestures, recognizing sign language, etc. In our SLAS, there will be two major components we want to design and implement: (i) a robust gesture recognition engine (GRE) that uses posture and motion information obtained from the hands and the fingers. (ii) a routine learning engine (RLE) that uses machine learning algorithms to learn motion patterns, actions, and to configure and control the devices accordingly and automatically. This project will involve spatial and temporal pattern recognition in the form of gesture recognition. Our developed algorithms will work on depth data captured by depth cameras deployed in the environment. Spatial and temporal patterns will be recognized using different classifiers that better suit to the specifics of the problem. Research will be performed on extracting discriminative spatial and temporal features, classifier training on the training sets, and joint training of the spatial and temoral classifiers. A further training of our SLAS will be performed on people with disabilities.Joint With Zet Technology and Physical Disabilities Foundation
Energy-Aware Mobile Gaming and Virtual Environments
Despite the increasing popularity of mobile devices and online games, there are certain shortcomings that create challenges for designing mobile games, which by nature are resource-demanding and/or real-time applications. The most important of these challenges are: (i) Limited energy: : mobile devices have limited energy supplies as most are battery powered. Thus, minimizing energy consumption while maximizing the game’s quality is a difficult problem. (ii) Limited bandwidth: mobile devices currently have more bandwidth limitations than wired PCs. While this will improve in the future, it is currently imposing difficulties for bandwidth-demanding games. In addition, using more bandwidth, even if available, will contribute to more energy consumption, pushing us back to challenge 1 above. (iii) High network delay due to the nature of wireless networks, geographically-distant users will experience a high communications delay, more so than in wired networks. This is a significant obstacle for highly-interactive and tightly-synchronous multiplayer games. (iv) Limited processing and display: while mobile devices are progressing rapidly, they are still limited in terms of processing power and graphics display, both of which adversely affect the quality of next generation systems (v) Heterogeneity: another layer of complexity is added to all of the above when considering that mobile devices are not all the same and have different computing and communications capabilities and/or use different networks with different capacities. It therefore becomes crucial for mobile games to be context and content aware and to adapt to specific configurations and variations of individual players The purpose and goal of this project is to study, design, and develop energy-aware mobile gaming technologies that will help reduce the amount of power consumption by games on mobile devices, leading to longer playing time and higher gaming quality for players.Principle Investigator: Shervin Shirmohammadi
Supported By TUBITAK under the Guest Scientist Support Program
TTelligent Labs: T2C2
TTelligent Labs aims to provide a scalable data analysis framework for a wide range of enterprise customers. The framework is provided as a service and runs purely in the TT compute cloud (T2C2). The T2C2 will be built and used for the project and also become a project deliverable. The customers use T2C2 to run complex data analysis jobs that are otherwise “difficult” to put together, time intensive to compute, and hard to scale using traditional siloed approaches (e.g., RDBMS, OLAP, Data Cubes). Furthermore, there is a vast amount of “non-standard” real-time and micro data (e.g., Twitter tweets, Facebook likes, Del.icio.us bookmarks, Linkedin group memberships, Pinterest pins, Google+’s and etc.) that you cannot possibly integrate all in one place. Customers need a parallel and immensely scalable set of adhoc analysis tools to crunch data to drive real time business insights. TTelligent Labs facilitates this insight extraction process. The main driver is to transform the business data analytics space by making business and data analysts get used to the following frame of thinking: What insight would I gain if I had full use of a 100-node compute cluster (C2) for an hour? What if one hour of this 100-node C2 would cost me only 50 TL? In the traditional approach, a business user first determines what question to ask, and then; IT would massage the underlying data to answer that query. In the new approach (T2C2), IT provides a platform that enables the creative discovery process. Then, the business user starts exploring with questions. This is the future of computing and doing business going forward.Principle Investigator: Ahmet Bulut
Mobile Robots P3-DX (aka Robodog)
In this project we want to develop a tracking system using Kinect sensor and Mobile Robots P3-DX. Some of the shortcomings of existing robot tracking systems are the ability to deal with obstacles and identifying the target. Our system will be able to track nearest human even under bad conditions (e.g. existence of obstacles and corners). The system will be implemented using ARIA and Microsoft Kinect SDK.
Optical Character Recognition
Machine Learning Lab is glad to announce the latest project: Optical Character Recognition! (OCR) Nowadays, we are working with the MNIST dataset to develop an Optical Character Recognition system.
Intelligent Scoping of Paid Search Campaigns using Advertiser & End User Provided Relevance-Feedback
We are planning to build an automated search campaign management tool for paid search advertising. Our goal is to increase sales while easing the campaign management for advertisers. Our approach is to unify the positive end-user feedback in the form of conversions with the negative feedback provided by advertisers for building a conversion model.Principle Investigator: Ahmet Bulut
Topic Strand: Analyzing Social Media Discussions w.r.t. Participation Patterns and Participant Profiles
Social media discussions exhibit rich temporal dynamics with respect to their occurrence as well as the profiles of the participants engaging in the discussions. For instance, traffic problem of İstanbul is a topic occasionally appearing in the social media with varying discussion density, where a wide range of people participate in the discussions. On the other hand, the discussions about the Kurdish problem have increased density during times coinciding with major terror events, political activities involving solution initiatives, etc. The participant profile of such discussions is quite different compared to those from the traffic discussions. These examples show that in social media platforms like Twitter, a topic’s discussion characteristics and temporal change of these characteristics over time show variety across topics. Despite these characteristics being dependent on the participant behaviors and human behavior being difficult to predict and model, the dynamics of the discussion, which is a result of the collective behavior of the participants, can be used to compare topic discussions with each other. This also makes it possible to cluster topic discussions based on their collective characteristics. Our goal is to analyze social media discussions, find similarities in between different discussions, cluster topics and do sentiment analysis using machine learning techniques.
Intelligent Regularization: Sensitivity-Steered Message Passing
Our goal in this project is to solve two important research problems: (i) finding a method to guide the perturbation of the inverse problem (ii) designing the message-passing algorithm so thatit exploits the underlying graphical model and is a GPU-friendly algorithm. The first research problemwill enable probabilistic inference algorithms that use the prior information about a solution in a moreintelligent way and perform better than the conventional MAP estimation algorithms. The second researchproblem will enable the use of these algorithms in high-complexity problems of computer vision via theirimplementation on GPUs. These two issues are discussed in more detail below: (i) Prior work in this area uses linear relaxation techniques to design scalable algorithms that work byexchanging messages and can achieve convergence under some predetermined constraints. Our approach is touse linear relaxation techniques not solely for satisfying the constraints of the problem that encapsulatesthe perturbation of the original problem but for inferring if a constraint is in agreement with the data.To be able to infer if a constraint is in agreement with the data we are going to use Bregman’salgorithm which applies the method of successive orthogonal projection in the primal-dual algorithmicscheme. The dual variables are the sensitivities of the original problem with respect to the constraintscorresponding to the dual variables. By using the dual variables, which are byproducts of the Bregman’salgorithm, the sensitivities of the problem with respect to each constraint will be learnt and the constraintwill be relaxed or removed accordingly as the Bregman iterations progress. By doing so, the assumptionsand the prior information about the solution will be applied intelligently according to how well theyagree with the data. (ii) Message passing algorithms converge to a solution by exchanging messages locally. The messages aresent to neighboring nodes/vertices as determined by the underlying graphical model of the original problem.Since message passing algorithms can do their updates asynchronously they can easily be implemented onGPUs, which are manycore computing hardwares. This is because in a asynchronous updating scheme eachnode can be mapped to a GPU thread without having to depend on other nodes/threads. This property willenable bigger faster execution on the GPU as compared to the CPU. Specifically, computer visionapplications, which will be our focus, the unknowns such as optical flow vectors, and disparityfields lie on a 2D grid as dictated by the problem. Hence, message passing nodes can be easily mappedto GPU threads. However, the real achievement will be to utilize the stream processing features of GPUcomputing. Our aim is to partition the nodes in the underlying graph into sets that can performcomputations in a parallel manner by using asynchronous update schemes. These partitions will be assignedto GPU streams that can work independently hence further increasing the speed-up achieved on the GPU.Our goal is to achieve a milestone in solving inverse problems. Since inverse problems are basicallyunderdetermined problems, some assumptions and prior information are utilized to improve the ill-definednessof the problem. We are going to infer from the data and the iterate if the constraints that encapsulatethese assumptions and prior information are correct or not by using Bregman’s algorithm. This will be abig improvement in the theory of solving inverse problems and optimization in general. It is importantto note that this is not parameter estimation or learning the graphical structure using training datasets. Our method is going to achieve the above described goal while doing modified Bregman iterationson the test data, not the the training data. In short, the original problem will be modified accordingto the problem data while actually solving the problem. Hence, the estimation problem will depend on theinput data. This is an innovative approach to solving inverse problems. And we believe this will be animportant milestone in solving inverse problems. Also, since we will implement the resultant algorithmson GPU, the algorithms can be used in solving high complexity problems of machine learning and computer vision.Principle Investigator: Tarik Arici
Supported By TUBITAK under the 3501 Young Researcher Career Support Program
Real-Time Video Analytics Engine Optimized for GPUs
Video surveillance systems are widely deployed to keep private and public spaces safe and secure.There are over 30 million cameras in United States only, shooting 4 billion hours of video footage a week.Currently, it requires significant human supervision to analyze the videos captured by surveillance cameras.Since it is not possible to analyze all the video data with eye inspection, most of it is stored and not processed.One approach to tackle this problem is to simplify the algorithms, but this would inevitably increase thefalse alarm and miss rates. Another approach would be to use more computing power. Programmable Graphics Processor Units (GPUs) have evolved into multi-threaded, many-core, highly parallel processors. However, to be able to take full advantage of the GPUs, the algorithms must be highly parallel. The objective of this project is to design and implement parallel video analysis algorithms optimized for the GPU.The state-of-the-art computer vision algorithms used for video analysis will be parallelized or newalgorithms optimized for GPUs will be designed if needed, without compromising the performance. Theoretical work on global optimization using message passing will be done so that the convergence is fast andresulting local minimum is satisfactory. The message passing algorithm will be used in optimization stagesrequired by most of the analytics algorithms. A metadata will be created, shared and used by the algorithmsto reduce the overall running time. An analytics engine will be designed and implemented to efficiently usethe results of this project, while achieving a (close to) real-time execution. Finally, this engine will betested on video footage obtained from Istanbul Police Department’s Information and Security System, which isa video surveillance system that monitors Istanbul’s streets, highways, and important districts with highcrime rates, accidents, and congestions.Principle Investigator: Tarik Arici
Supported By European Union under the 7th Framework Programme