Projects (CASLAB)

The proposed project has two main goals. The first one is improving the Turkish natural language processing resources
developed mainly by the researchers in the project team based on the current technology. The second goal is compiling all these
resources in an open platform and thus contributing to new research. In the scope of the university, there are plenty of resources
developed by the researchers working on natural language processing. In the project, these resources will be collected, will be
improved by the support of linguistics researchers and a natural language processing platform will be built. The platform will be
open-sourced and will be open to development. In this respect, resources required for Turkish natural language processing will be
opened to researchers.

Sentiment analysis is the process of extracting the sentiment (positive, negative, neutral) in texts using natural language processing and machine learning techniques. This process has five components: the sentiment, person owing the sentiment, time period of the sentiment, the object and the aspect of the object related to the sentiment. There is no sentiment analysis work for Turkish that takes these five parts into account. In this project, we will form a sentiment analysis framework for Turkish. The works conducted for foreign languages cannot be adapted to Turkish easily. Moreover, most of the sentiment analysis studies for Turkish employ supervised methods; they do not use unsupervised or semi-supervised methods. This poses problems when supervised data are limited. Also, Turkish studies usually give rise to less satisfactory results compared to other languages. Other issues are forming sentiment vectors and building an aspect-based sentiment analysis system. In this project, a framework for Turkish sentiment analysis that incorporates polarity, its score, related object, its time, and owing person. For this purpose, neural networks and unsupervised/semi-supervised methods will be used. Finally, domain-specific sentiment lexicons will be built.

Bu projede, Türkçe gibi eklemeli dillerde geniş sözcük dağarcıklı sürekli konuşma tanıma (large vocabulary continuous speech recognition – LVCSR) sistemleri için kullanılacak yeni bir dil modelinin geliştirilmesi amaçlanmaktadır. Bilindiği gibi, eklemeli dillerde sınırsız sayıda kelime üretilebilmesi, konuşma tanıma sistemlerinde dil modeli oluşturmada zorluklara neden olmaktadır ve iyi bir dil modelinin eksikliği bu sistemlerin etkinliğini önemli ölçüde etkilemektedir. İngilizce gibi nispeten eklemeli olmayan dillerde konuşma tanıma sistemlerinin başarıyla geliştirilmiş olmasının ve Türkçe gibi dillerde henüz aynı başarıya ulaşılamamasının en önemli nedenlerinden birisi, etkin bir dil modelinin eksikliğidir. Konuşma tanıma sistemlerinde yaygın olarak n-birimli (n-gram) dil modeli kullanılmaktadır. Bu model, dili istatistiki olarak modellemeye çalışmaktadır. Geniş bir metin havuzundan (corpus) sözcüklerin birbiri ardına gelme sıklıklarını göz önüne alarak oluşturulan model, sözcük dizilerinin olasılıklarını hesaplamakta kullanılmaktadır. Türkçe’de diğer bir sorun, sözcüklerin cümle içinde diziliminin nispeten serbest olmasıdır. Bu serbestlik sorunu Türkçe’nin eklemeli bir dil olması ile birleştiğinde, sözcük bazında basit bir n-birimli dil modelinin konuşma tanıma sistemlerinde etkinliğini azaltmaktadır. Bu çalışma ile Türkçe’nin morfotaktik (morphotactic) (morfların dizilim kuralları) bilgisini n-birimli bir dil modeli ile birleştirerek etkin bir dil modelinin oluşturulması hedeflenmektedir. Böylece Türkçe için birçok uygulama alanına sahip geniş sözcük dağarcıklı konuşma tanıma sistemleri geliştirilebilecektir.

Bu projede, spam e-posta mesajlarının önlenmesine yönelik olarak Türkçe için spam-önler filtreleme metotları geliştirilecektir. Günümüzde spam mesajlar tüm e-posta mesajlarının %10’unu oluşturmaktadır ve kullanıcılar açısından önemli zaman kayıplarına neden olmaktadır. İngilizce gibi yaygın diller için filtreleme algoritmaları mevcuttur, fakat Türkçe mesajlar için henüz böyle bir çalışma yapılmamıştır. Bu tür bir çalışmada Türkçe’nin karmaşık morfolojik yapısının gözönüne alınması gerekmektedir. Bu projede geliştirilecek olan metotlar dinamik olacaktır ve yapay sinir ağları ile Bayesian ağları tekniklerine dayanacaktır. Ortaya konulacak olan algoritmaların iki temel bileşeni içereceği düşünülmektedir: Mesaj içeriklerinin morfolojik analizini yapacak bir morfoloji modülü ve mesajları normal ve spam olarak sınıflandıracak bir öğrenme modülü.

Small Unmanned Aerial Vehicles (UAVs) are used in many areas including search & rescue applications, geographic mapping, disaster management, precision agriculture, wildlife monitoring, shipping, delivery, and aerial photography. However, there are also undesired/malicious usages of drones such as: terrorist attacks, surveillance for restricted/critical zones, and smuggling across borders/prisons. Therefore, there are regulations for drone usage to identify these threats.

The CONNECT project aims to develop a localization system to prevent undesired/malicious use of drones. In the literature, there are proposals and solutions focusing on single sensor type. CONNECT will be a multi-sensor platform including sound, radio frequency (RF), and optic sensors to overcome the weaknesses of each sensing mechanism. These kinds of measurement systems with multiple types of sensors are called multimodal acquisition systems. CONNECT will consist of three sensing stations and each of them will be equipped with RF-based, acoustic, and optical sensing devices. For complex phenomena like drone localization, it is not expected to have a single modality that provides complete and robust knowledge. Therefore, each modality introduces a diversity that can be exploited. In addition to drone detection and localization, CONNECT will also aim to localize the remote controller units, If there is a communication link between the drone and the control unit.

Model-driven approaches require realistic models of the underlying processes and they always come with some assumptions, which may not be plausible. For complex and multimodal systems, model-driven approaches may not be the best choice for revealing the relationships between modalities. On the other hand, Machine Learning (ML) algorithms and artificial neural networks architectures are frequently used to perform detection, classification, and tracking within the data-driven framework. Data-driven approaches have been successful in human motion tracking, object detection, astrophysics, audiovision, and more. By focusing on data-driven approaches, we plan to reduce the number of assumptions and external inputs. We will compare datadriven and model-driven approaches within the modalities and we aim to improve the localization accuracy of the CONNECT by incorporating ML-based approaches to the existing techniques within the modalities and the fusion 

This project focuses on a recent challenge in deep learning and its potential effects on healthcare systems that utilize deep models. Deep networks have demonstrated a strong capability in detecting, classifying, and recognizing patterns in various domains. However, a recent discovery showed that convolutional neural networks are vulnerable to specifically crafted perturbations, also known as adversarial perturbations. One of the

state-of-the-art adversarial defense approaches is adversarial training. However, adversarial training suffers from overfitting. The objectives of this project are two-fold. First, an alternative adversarial defense approach will

be designed to address the poor generalization performance of adversarial training. Secondly, potential effects of adversarial attacks in the healthcare domain will be investigated and the proposed defense technique will be

extended to healthcare applications. It has been shown that adversarial attacks on medical images alter clinical decision-making. In addition to medical images, Electronic Health Records (EHRs) are commonly utilized to

train deep models in healthcare. EHRs have different characteristics than images, such as heterogeneity and temporal dependence, which can be modeled using Recurrent Neural Networks (RNNs). Adversarial training for

RNN models have not been studied thoroughly in the literature. In summary, an adversarial training approach, which is more generalizable and adoptable by different deep architectures, is the main result to be achieved in this project.

Abstraction and abstract reasoning are among the most essential characteristics of high-level intelligence that distinguishes humans from other animals. High-level cognitive skills can only be achieved through abstract concepts, symbols representing these concepts, and rules that express relationships between symbols. This project aims to self-discover abstract concepts, symbols, and rules that allow complex reasoning by the robot. If the robots can achieve such abstract reasoning on their own, they can perform new tasks in completely novel environments by updating their cognitive skills, or by discovering new symbols and rules. If the objectives of this project are achieved, scientific foundations will be laid for robotic systems that learn life-long lasting symbols and rules through self-interacting with the environment and express various sensory-motor and cognitive tasks in a single framework.

In this project, a novel wearable exoskeleton with flexible clothing will be developed for the use of paraplegic persons who have lost their lower extremity motor functions due to low back pain, paralysis and similar disturbances. This prototype will be designed as a rigid-link exoskeleton that is actuated via series-elastic actuators. The user’s physical state will be observed using soft elements attached to a wearable sensorized clothing, which will be used in conjunction with the exoskeleton. Trajectory planning will be explored by a human-to-robot skill transfer problem to minimize the metabolic cost. Stability and balance analyses will run, and environmental factors will be estimated to provide safe walking support to the human using a hierarchical control structure.

Project description

Cybersecurity is vital to any person or entity, from consumer to government, involved in conveying information. The key lies in being able to detect attacks and react quickly and efficiently by launching appropriate countermeasures. The ITEA 2 project ADAX has delivered a set of key innovations improving prevention, detection, decision support, countermeasure enforcement and knowledge management to support security operation on complex and critical IT infrastructures.

 

DETAILS:

http://adax.boun.edu.tr/

İşaret dili, işitme engellilerin ana dilidir. İşaret dili, mesajı iletmek için pek çok görsel öğeden yararlanır: El şekli ve yörüngesinden oluşan jestler (işmarlar); vücut pozu, ve yüz ifadeleri. Her kültürel grubun ayrı bir işaret dili vardır ve Türk işaret dili (TİD), Türkiye’deki işitme engellilerin ana dilidir. TİD’den Türkçe yazılı ve sözlü dile tercüme ve bunun tersi, ancak her iki dili de bilen ve işitme engeli olmayan tercümanlar yoluyla yapılabilmektedir ve pahalıdır. Örneğin, ancak sınırlı sayıda televizyon programı, TİD ile simültane tercüme edilerek yayınlanmaktadır. İşaret dilinin otomatik tanınması, bilgisayarla görmede çözülmemiş zor problemlerden birisidir. İşaret dili tanıma teknolojisinin gelişmesi, işitme engelli bireylerin topluma katılmasını kolaylaştıracak ve hızlandıracaktır. Öte yandan, işaret dili tanıma, bilgisayarla görme teknolojisinin gelişmesi için öncü rol yapmaktadır; ve bu alandaki gelişmeler insan- bilgisayar etkileşiminde de kullanılmaktadır.

Bilgisayarla görmede derin öğrenme yaklaşımı, son yıllarda, nesne tanımada büyük başarı kazanarak, en tercih edilen yöntem haline gelmiştir. Bu gelişmeyi hazırlayan etkenlerden birisi, çok büyük ölçekli imge veri tabanlarının toplanması ve içeriğinin işaretlenmesi olmuştur. İşaret dili alanında, benzer şekilde, her dilin işaret diline ait geniş veri setleri toplanmaktadır. Bu projede amacımız, toplanan bu veri tabanlarını kullanarak işaret dili için özelleşmiş bir derin sinir ağı eğitmektir. Derin sinir ağlarının eğitiminde, başka bir alanda eğitilen bir ağın, hedeflenen alanda kullanılması için transfer öğrenmesi teknikleri kullanılır. Eğitilen derin ağ, değişik işaret dillerinin betimlenmesinde kullanılabilecek ortak bir gösterim verecektir; ancak başarımının artırılması için, hedef alandaki veriyle tekrar eğitilmesi gerekir. Bu amaçla, TİD’de toplanmış kaynaklarda transfer öğrenmesi yöntemlerini kullanacağız.

Yapılan araştırmalar sonucunda, işaret dilinde yapılmış ayrık bir işaretin tanınması ve bir video içinde işaret dili ile sorgu yapılması için sistemler geliştirilecektir. Bu sistemlerle, işaret dili bilmeyen bir kişinin, gördüğü bir işaretin anlamını öğrenmesi, ve bir işitme engelli kişinin, işaret dili ile videolarda sorgu yapması mümkün olacaktır. Aynı zamanda, işaret dili ile işaretlenmiş sesli videolarda yapılan sorgularla, videodan sese ve yazıya tercüme yapabilen TİD sözlüğü geliştirilmiş olacaktır.

https://www.cmpe.boun.edu.tr/~emre/projects.html#imagine-cog

In this research project, we aim to design and implement an effective robotic system that can infer others' goals from their incomplete action executions, and that can help others achieving these goals. Our approach is inspired from helping behavior observed in infants; exploits robot's own sensorimotor control and affordance detection mechanisms in understanding demonstrators' actions and goals; and has similarities with human brain processing related to control and understanding of motor programs. Our system will follow a developmental progress similar to infants, whose performance in inferring goals of others' actions is closely linked to development of their own sensorimotor skills. At the end of this project, we plan to verify whether our developmental goal inference and helping strategy is effective or not through human-robot interface experiments using upper body Baxter robot in different tasks.

https://www.cmpe.boun.edu.tr/~emre/projects.html#afford

Bu proje ile, ortamın robota sunduğu sağlarlıkları (affordances) öğrenip modelleyerek sağlarlıklar ve sensör geribildirimleri ile desteklenen gelişmiş bir manipülasyon beceri sistemini kurmayı hedeflemekteyiz. Bu tip ortamlardaki `tutmak', `taşımak' ve `yerleştirmek' gibi eylemler tipik oldukları için, hareketleri, gösterim yolu ile öğrenme (learning by demonstration) ile robota aktarmayı planlamaktayız. Bu yolla yarı-yapısal ortamlar için gerekli manipülasyon becerilerini öğrendikten sonra, robot, ortamın sunduğu görsel ve diğer sağlarlıkların, bu becerilerin yürütülmesini nasıl etkilediğini öğrenmelidir.

Introduction:

The European manufacturing landscape is in urgent need of change. In 2014, manufacturing represented about 16% of the EU GDP, more than 80% of EU total exports, 80% of private Research & Development expenditure, and employed 30 million people. However, the financial crisis has heavily hit the sector, combining its negative effect with ongoing globalisation and technological innovation negative externalities. Together these factors have culminated in the loss of over 3.8 million jobs.

Reinvigorating the manufacturing sector is a complex task that requires transformation at the heart of production processes and models upon which our industrial society has been built. Inroads in generating this scale of shift in manufacturing have been made in areas conducive to open technologies in the fields of digital fabrication and craftsmanship – where the manufacturing sector has come into contact with the emerging social-technology based Maker movement, innovation and growth emerges. This drives the democratisation of production, turning manufacturing into a participatory, collaborative, and open process in which all agents share risks and benefits and, ultimately, increase the value of production. 

More information here

Call: H2020 Collective Awareness Platforms for Sustainability and Social Innovation

Start Date: 20 June 2016 (24 months)

Project page: http://openmaker.eu/

The main aim of this project is to discover situations like verbal aggression, harrassment, and cheating in games over social media, and to provide a tool for automatically evaluating player complaints in terms of importance and veracity. Such a system will reduce the moderation costs of gaming companies, as well as constitute a tool for analysis of gamer behavior.

-

The principal goal of this project is the development of a depth based sign language recognition prototype and related demo applications, which hearing impaired users will make use of to aid in their communications with healthcare and finance professionals in their daily tasks. 

According to the 2000 DIE census, there are 109.000 people with total hearing disablity in Turkey. In their daily routines, the hearing impaired , are forced to use either written materials or the aid of an accompanying Turkish sign language interpreter to establish basic communication since they are unable to use speech as a medium of communication. The staggeringly low literacy rate among the hearing impaired greatly reduces the integration of this population, thus creating both a social and an economic disadvantage.

Sign languages are the main communication medium of the hearing impaired people. Sign languages convey meaning through hand movements, facial gestures and upper body postures.These languages differ from country to country. The Turkish Sign Language (TİD) is used by hearing impaired people of Turkish origin to convey their meaning through combinations of hand movements, facial gestures and upper body postures unique to it. In 2005, as part of the European Union integration effort, the Turkish Social Services law made it mandatory for all governmental organizations and offices to employ a TİD interpreter. In 2006, the use of TİD and the task of training TİD interpreters were introduced through regulations into the Turkish Social Services law.

These legal requirements constitute one of the emergence factors of this project. Since it is neither practicable nor economically feasible to train and employ TİD interpreters for every government office, practical solutions like establishing sign language translation call centers that employ TİD interpreters are considered. The ideal solution to such a dilemma is the recognition and translation of sign languages to speech through the use of technology. Such a system would diminish the need for sign language interpreters, while fully integrating the hearing impaired. However, the technology needed for such a system is beyond the current state of the art. Today’s technology only allows for sign-to-speech translation systems with extremely limited corpora that are heavily user dependent. The goal of this project is the realization of a sign language recognition system that performs successful recognition over an extended vocabulary and is signer independent.

For over half a century, Netas has presented communications solutions to both overnmental and private companies. The WEBRTC platform that is being developed by NETAS allows for two way real time text, audio and video communication between Internet browsers and classical networks. It is one of many solutions, which the company hopes to integrate with sign language recognition and accesibilty modules in the future.

One of the main goals of this project is to develop a computer vision based system that will recognize a chosen corpus of significant TİD gestures. The aim of the system will be to present notification and guidance services in healthcare and finance domains. Taking the market potential of the project into account, some sample application domains for the project are as follows: 

1- Hospital Information Desk: Providing guidance, direction and appointment services to hearing impaired users.
2- Bank: Providing guidance and service related information to hearing impaired users.

The university group of the project is well versed in the theoretical and application side of the sign language recognition domain. (Detailed information and full list of publications is available at: http://www.cmpe.boun.edu.tr/pilab/doku.php?id=research:sign_language_rec...) The research can now be considered to have reached a prototyping stage. With this project, a prototype will be developed which can be turned into products in the future.
The scientific goal of this system is to realize a state of the art TİD recognition and human computer interaction software. The software will make use of transfer learning and domain adaptation methods to enhance its sign vocabulary and establish increased signer independence. The TİD depth video dataset, which we aim to collect in this project will be a first in this field considering its size and scope and will serve as a valuable resource to
researchers working in this field.

Social software is defined as software that enables its users to interact socially on the Web through organization structures such as a social network. The use of social software spans casual friendship networks to complex process management. The number of users of social software of various sorts is huge, for example Facebook.com by itself has more than one billion users. Users of these systems typically share a vast amount of content. The content that is being shared is always targeted for some of the users of the system but not for all of them. Every user has the right to demand that her privacy is preserved by not showing the content to those individuals for whom the content was not meant for. However, everyday more and more interesting cases of privacy breaches are taking place. A privacy breach by itself is not desired but it can have further effects. For example, if a private credit card information is shared with unknown people, it can result in serious security problems. Or, an invention that needs to be secret in a company is mistakenly shared with a competitor company, the invention details may leak. For these reasons, it is extremely important to preserve users' privacy in social software. To enable this, we need intelligent software that can keep track of users' privacy requirements, check if these requirements can be met by the system, and lead the user to take appropriate action as needed. For example, the software can suggest the user not to share a certain piece of information if the social software cannot guarantee its privacy or the software can signal that an important content has been shared with undesirable individuals. Accordingly, the aim of this project is to design and develop the necessary components and techniques for an intelligent software that will manage a user's privacy on her behalf.

The main aim of this project is to discover situations like verbal aggression, harrassment, and cheating in games over social media, and to provide a tool for automatically evaluating player complaints in terms of importance and veracity. Such a system will reduce the moderation costs of gaming companies, as well as constitute a tool for analysis of gamer behavior.