Thesis defense

thesis defense

Thesis defense - Ali Kaan Sungur

Graduate School of Informatics /Multimedia Informatics

In partial fulfillment of the requirements for the degree of Master of Science Ali Kaan Sungur will defend his thesis.

Title: HIERARCHICAL TEMPORAL MEMORY BASED AUTONOMOUS AGENT FOR PARTIALLY OBSERVABLE VIDEO GAME ENVIRONMENTS

Date: 15th August2017

Time: 11:00 AM

Place: A-212

Thesis Abstract : Believable non-player characters (NPC) can have a profound impact on the experience that a video game provides. This thesis presents an online, unsupervised and lifelong learning autonomous agent that the player can interact with. It has an architecture that is based on a combination of Hierarchical Temporal Memory and Temporal Difference Learning Lambda with the guidance of neurobiological research. The agent has a visual sensor with online data stream. Input from the sensor is fed to the architecture in order to model the surrounding environment. The goal of the agent is to learn rewarding sequences of behavior based on the stimulation it receives caused by its own actions. It navigates in a procedurally generated three dimensional environment and is in a continuous learning state adapting the synapses of its neural connectum. In addition, the architecture is capable of being stored and loaded at any point allowing for persistent learning through multiple simulation sessions. The study presents the characteristics of the architecture on a video game related learning task. Data collected from the experiments with varying parameters are compared along with the runtime and serialization performance. The proposed methodology results in an autonomous NPC that can learn rewarding behaviors without any supervision. Moreover, it is also capable of learning specific action sequences via player guidance. End result is a promising and novel NPC architecture that is also fairly open to incremental improvements through the relevant neurobiological studies and the advancements on the theory of Hierarchical Temporal Memory.

Announcement Category

Thesis defense - Ehsan Zabardast

Graduate School of Informatics /Health Informatics

In partial fulfillment of the requirements for the degree of Master of Science Ehsan Zabardast will defend his thesis.

Title : PREDICTION OF SURGICAL OPERATION DURATIONS USING SUPERVISED MACHINE LEARNING TECHNIQUES

Date: 11th August 2017

Time: 15:00

Place: A-212

Thesis Abstract :There’s an ever increasing number of patients referred to healthcare facilities and hospitals. The healthcare facilities have two main options to deal with this situation. They have to either employ and acquire more resources or they should use the existing staff and resources more efficiently and effectively. The first option is not always feasible due to the fact that the healthcare facilities have limitations on both the staff they can employ and the resources they can acquire. Given the fact that these resources are expensive and extra resources provide diminishing returns, it is important to make the best use of resources available. Operating rooms and surgeons are the most expensive and scarce resources in hospitals; so it is crucial to optimize their performance and avoid under and over utilized operating rooms. The aim of this study is to employ supervised machine learning techniques and probabilistic graphical models to predict the duration of surgical operations using historical data. We have used a wide spectrum of different models ranging from regression methods, classification methods, and Bayesian Networks to predict the surgical operation durations. The models built based on Bayesian Networks, in general, produce more accurate results with lower errors. Naive Bayes, however, outperforms the other Bayesian-Network based models with an average accuracy of 66.9% and root mean square error of 998 seconds (16.6 minutes) from the true duration of the operation. Provided with accurate estimation of surgical operation durations, it is possible to build optimization models to utilize healthcare facility resources. This allows healthcare facilities’ managers to create tactical (medium term) plans and to increase efficient utilization of operating rooms and surgeons.

Announcement Category

Thesis defense - Mahdieh Farzin Asanjan

Graduate School of Informatics /Health Informatics

In partial fulfillment of the requirements for the degree of Master of Science Mahdieh Farzin Asanjan will defend his thesis.

Title: SEMI-AUTOMATIC SEGMENTATION OF MITOCHONDRIA ON TRANSMISSION ELECTRON MICROSCOPY IMAGES USING LIVE-WIRE AND SURFACE DRAGGING METHODS

Date: 16th June 2017

Time: 10:30 AM

Place: A-108

Thesis Abstract : Semi-automatic segmentation of mitochondria Segmentation is an integral part of image processing. Contrary to many technical applications the design of fully automated segmentation routines is extremely challenging in the medical context because of the large biological variation. Even if automatic routines do work in normal subjects, they typically fail in pathologic cases, which are often more interesting from a clinical point of view. Segmentation of mitochondria in medical images is essential for studying mitochondrial morphology and computer aided analysis and diagnosis. Since using an automatic segmentation method may leads to the least flexibility and also using the manual methods needs a considerable amount of human effort, automatic detection and segmentation of mitochondria with user interaction is necessary in order to facilitate the analysis of large 3D data sets and user interaction introduces a subjective element to image processing and analysis. So segmentation methods using human interaction to initialize the algorithms can be more helpful. In new method I am trying to make possible to operator interaction although it may be used in a minority of cases, only. And this will help to reduce the failure rates. Two principle modes of user interaction can be distinguished. In the first mode the user interactively selects a region or volume of interest (ROI or VOI) in which subsequently an automated operation is performed. A typical well-known example is the selection of a seed point to start a region growing process. In contrast, the second mode is iterative and requires extended user interaction, for example, an interactive change of a contour.

Announcement Category

Thesis defense - Özgen Demirkaplan

Graduate School of Informatics /Cognitive Science

In partial fulfillment of the requirements for the degree of Master of Science Özgen Demirkaplan will defend his thesis.

Title: EFFECTS OF VOICE FAMILIARITY ON AUDITORY DISTANCE PERCEPTION

Date: 20th June 2017

Time: 14:00 PM

Place: A-108

Thesis Abstract : Alongside the audible content of sound sources, human auditory system also provides the necessary cues enabling humans to perceive the direction and determine the range of sound sources, thereby allowing the recognition of spatial context. While the body of knowledge on the directional localisation of sound sources is extensive, less is known about the auditory and cognitive cues that allow distance perception. One such cognitive cue is auditory familiarity: Humans can judge the distances of familiar sound sources more accurately. In other words, the level of familiarity with a sound modulates the distance perception accuracy. However, auditory familiarity is not a rigorously defined property of sound sources. No objective or subjective scale exists that can reliably be used to assess familiarity of a subject with a given sound source. This makes a rigorous study of the effects of auditory familiarity on distance perception impossible. In order to assess this aspect of distance perception, this thesis aims 1) to develop an objective scale for auditory familiarity based on behavioural experiment , and 2) to investigate the effects of auditory familiarity on distance perception using source signals whose auditory familiarity can be compared by using the developed familiarity scale. This study investigates that whether the auditory distance perception can be enhanced by a sound source coming from a personally familiar voice.

Announcement Category

Thesis defense - Umut Demirel

Graduate School of Informatics /Game Technologies 

In partial fulfillment of the requirements for the degree of Master of Science Umut Demirel will defend his thesis.

Title: CREATING A GENERIC HAND AND FINGER GESTURE RECOGNIZER BY USING FOREARM MUSCLE ACTIVITY SIGNALS

Date: 01th August 2017

Time: 13:30 PM

Place: A-212

Thesis Abstract : Hand and finger gestures are the most natural way of communication without speaking. Therefore using hand and finger gestures as inputs is widely used in human-computer interaction applications. There are variety of applications using gestures as inputs such as: sign language recognition, robot controlling, medical device controlling and video game controlling. This thesis deals with creating a generic hand and finger gesture recognizer by processing muscle activity signals. For data collection, MYO from Thalmic Labs, a portable armband having 8 channel EMG sensors and IMU, is used. Normally, EMG sensors should directly be placed on muscles for each user. However, because MYO is a generic device for anyone to use, EMG sensor placements will differ from user to user. As a result, the recognizer will be calibration based, meaning that it is person and session dependent. Gestures are chosen from the set of expressive gestures used by classical orchestra conductors.13 different hand and finger gestures and 1 rest gesture is performed 5 times, each for 3 seconds by 20 subjects in separate sessions. While preprocessing the collected data, not only the time domain is used by applying filters to raw data, but also the frequency domain is used by representing each gesture with circular harmonic coefficients. Neural Network is trained, validated and tested to recognize 14 different gesture patterns. Cross validation method is performed with 5 data sets for each test subjects. 3 sets are used for training, 1 set is used for validation and 1 set is used for testing. 20 combinations of data sets are fed to 20 different neural network for the cross validation. The proposed generic gesture recognizer system is tested with one of the best Turkish classical orchestra conductor, Orhun Orhon. Before the performance, his forearm EMG data is collected while he performs his own gestures. During the performance, data is continued to be collected and real-time accuracy is tested.

Announcement Category

Thesis defense - Erdem Can Irmak

Graduate School of Informatics /Game Technologies 

In partial fulfillment of the requirements for the degree of Master of Science Erdem Can Irmak will defend his thesis.

Title: 3D INDIRECT SHAPE ANALYSIS BASED ON HAND-OBJECT INTERACTION

Date: 13th June 2017

Time: 11:00 AM

Place: A-212

Thesis Abstract : In this thesis, a novel 3D indirect shape analysis method is presented which successfully retrieves 3D shapes based on the hand-object interaction. In the first part of the study, the human hand information is processed and transferred to the virtual environment by Leap Motion Controller. Position and rotation of the hand, the angle of the finger joints are used for this part in our method. Also, in this approach, a new type of feature is implemented which is called interaction point. These points are placed on the digital hand model and indicate whether the hand touches the 3D shape or not. In the second part, every 3D shape is represented by applying hand features to the Support Vector Machine. Experiments validate that Support Vector Machine results are usable for retrieval of 3D shapes. Moreover, in order to compare the retrieval results of this method, another interaction based indirect analysis method is implemented using Data Gloves. These two experiment results show which features are significant for retrieving the corresponding 3D shapes. The main contribution of this thesis is threefold: i) Noisy and/or deficient 3d shapes can be retrieved ii) The retrieval is not affected by the alignment of shape iii) Performance of the method is independent of the polygon count of the shape.

Announcement Category

Pages

Subscribe to RSS - Thesis defense

©2017 METU Graduate School of Informatics. All rights reserved.
Design: CC-IG