Date: 11/24/09

The objective of test evaluation is to dichotomize between bad chips and good chips. In other words, testing is applied as a way of filtering. Traditional test evaluation techniques include test-metric and fault model evaluation that rely on “large” and “time-consuming” silicon-based test experiments. In these techniques, new test method must be generated and incorporated into existing ones which incurs high expenses. In this talk a test-metric evaluation methodology that uses the “readily-available test measurement in chip-failure log files” was presented. The metrics used in this methodology, include “stuck-at”, “bridge fault model”, “gate exhaustive, “N-detect”, and “physically-aware-N-detect”. To validate the methodology, three types of chips were used, LSI test chip, NVIDIA GPU, and IBM ASIC. In each test case a combination of metrics were used. The result showed that the method is valid as its coverage in all of the test cases were high.

This talk was technical and the speaker did not provide enough background information. Thus, it was quite hard to understand the innovative of the methodology for people, like me, who does not have any background in this area.

Date: 12/01/09

In this talk an overview of research conducting by the speaker and his group was provided. The research focuses on determining human activities and feelings via sensors through learning techniques. The applications of this research are various, and it can mainly be used in human health and intelligent user interface. In human health, it can be used for depression assistant, deception detection, and health monitoring with wearable accelerometers.

The objective is to analyze human’s behavior and feelings by tracking facial behavior and movement (such as motion of lips, eyes, and chins), as well as body motions. For instance, in deception detection application, by analyzing facial behavior, it can be determine if the person is telling lie.

In this approach, first, user’s facial and body movement were collected and given to the developed learning algorithm to learn user’s behavior. Once the algorithm is trained, it can predict user’s behavior and feeling through his/her facial and body movement.

The challenge is collecting data in very fast movements, identifying noisy movements from signal ones, and distinguishing between combined movements (e.g., when both chins and head are moving).

To learn from the collected data two approaches were used, metric learning and joint classification and localization. In the metric learning technique three components are considered for facial movement and it is assumed that every facial movement is a linear combination of these three components. In the joint classification and localization, several data of two known types (e.g., depressed and non-depressed people) were given in order to identify the differences between the facial characteristics of them.

In my opinion, this research is very challenging, interesting, and its application especially the human health one is meaningful. The disadvantage of this talk was that the detail of the algorithm and the validation strategy were not elaborated.

Date: 11/16/2009

Location: Room 722, Benedum Hall

In this talk three concepts, administration, management, and leadership were discussed. Web definitions for each of the concepts, especially Wikipedia, are mostly confusing and do not clarify the difference among them. In some people’s perspective, administration, management, and leadership are the same, whereas in some others’ opinion they are totally different concepts. However, administration, management, and leadership are neither the same, nor irrelevant. According to the lecturer, administrator “performs duties that enhance organization’s operation”.  Manager is one who also does administration but, his/her duties are not limited to administration tasks.

Leader, on the other hand, is not a manager. There are major differences between these two, such as: manager is powered by position, whereas leader is powered by influence. Therefore, an employee can be even the leader of his/her own manager. Unlike manager who does administration tasks, leader must be innovative. Manager always asks how and when to do things, while leader asks what and why to do things. Despite leader who focuses on people, manager focuses on system. Finally, manager does things right, whereas leader does right things.

Leader is accountable, optimistic, and cooperative. Leader has vision, goal, and plan for reaching the goal. He/She clarifies his/her goals and plans by communicating with others. Leader is passionate about what he/she does. Leader’s actions “inspire others to dream more, learn more, do more, and become more”. Unfortunately, in schools such as MBA, students are taught how to become a good manager and administrator rather than a good leader.

In my opinion, although this talk was not research oriented, it was very helpful in sense of understanding how to become a good leader. Because in my perspective taking leadership is an important factor for becoming successful in any career, as well as in communicating with people. The only aspect that I disagree with the lecturer is that he mentioned in order to get promotion you need to be a good manager rather than a good leader. However, I think in order to get promotion being a good manager is required but not enough. A manager who communicates well, cares about people, takes leadership, and dedicated is more likely to get promotion rather than a manager who only cares about the system and rules.

Date: 11/11/2009

Location: NSH 1507 (CMU)

In this talk, the concept of eye movement for identifying the context was presented. Eyes are involved in every of our tasks including visual, physical, and cognitive. In each task, their movement and gesture are usually different than in the other task. The main approaches for recording eye movement include electro oculography (EOG) and ECG. The main difference between EOG and ECG is that in ECG relative movement can also be measured. In this research, a wearable device which can record eye movements using EOG technique was used. An algorithm was developed to determine the context. In this algorithm three characteristics for eyes were used, soccode, blink, and fixed. This algorithm was implemented and validated with 8 participants. 5 tasks (reading, writing, browsing, copying, no-activity) were assigned randomly to each participant. The experiment showed that eye movement analysis can be used for recognizing the activity (context). Another experiment was done to assess memory recall process by analyzing eye movements. In this experiment 12 unfamiliar and 20 familiar faces were shown to 6 participants. The experiment also showed that memory recall process effects the eye movement. Applications of tracking eye movement for identifying activities are various. Determining level of user’s interruptability and task engagements, and reading assistants are examples of such applications.

Despite the innovations of this research, it has some major shortcomings. Firstly, the wearable device used in this research is obtrusive and most probably would not be accepted by users to wear. Secondly and more importantly, the context discussed in this research merely depends on the type of activity, whereas context is more than just an activity and can also include time, location, purpose of the activity, and user’s mood. With respect to time, there are differences between reading, writing, watching, or browsing activities while the user is in an outdoor environment with sun shining, or in dark environment were light is not sufficient. With regard to location, the level of concentration, hence, eye movement is different when doing any of these activities in a noisy environment versus a quiet place. Furthermore, there are definitely differences in any of these activities when the context of the activity itself changes. For instance, reading a research paper, a friend’s email for lunch invitation and a friend’s email reporting another friend’s death, news (political, social, economical), and fun magazines are entirely different contexts. Finally, user’s overall mood and health, e.g., sleepy, fresh and energetic, stressful, upset, or sick affects the eye movement.

Date: 11/07/2009

Location: University of Pittsburgh, Scaife Hall, Lecture Room 2

In this workshop different topics including planning and establishing a new class, giving effective presentation, and grading were discussed.

Preparation:

To prepare a course, several steps are involved:

1)      Doing research

2)      Formulating topics

3)      Identifying course objective

4)      Selecting textbook

5)      Developing schedule

6)      Designing assignments

7)      Writing description

8)      Developing syllabus

9)      Preparing lectures

Doing research- The first step of course preparation is doing research on the course and the students who would like to take the course. With respect to course, the purpose and the level (graduate or undergraduate) of the course, number of credits, and its prerequisites, and with respect to students, number of the students who are taking the course, their background and expectation from the course, and their demographic information need to be explored.

Formulating topics- In this step, the topics covered in the course must be identified. To do this, in addition to using information gathered in the first step, one needs to know what are the prerequisites of the course and which courses are dependent on this course.

Identifying course objectives- Course objective must be clarified, so that students know what they should expect from the course.

Selecting textbook- Selecting a textbook requires investigating through the content of different books and their supporting materials (such as power point slides). While the textbook must be up-to-dated, its price should also be reasonable.

Developing schedule- Schedule is very important as it shows which topics are covered in each lecture, what the due dates of projects/assignments are, which days are holidays, and when exams’ dates are.

Designing assignments- Assignments must render the course’s objectives. While assignments should be challenges, they must be doable.

Writing description- one must make sure that described the course clearly and attractively. If the course description is not attractive or clear enough, students either do not take the course, or take it while expecting something else from the course.

Classroom performance:

To enhance the teaching performance, the instructor must make sure that he/she is both knowledgeable and enthusiastic enough to teach the course. If the instructor does not have enough knowledge to teach the course, he/she will not be able to deal with student questions. In addition, if the instructor does not interested into the topic, he/she cannot motivate students to listen to the lectures, study, and do the assignments.

The lecture should focus on generalized skills rather than unimportant details such as dates and numbers. To present the lecture more effectively so that students do not easily forget the content, it is important to tell stories and bring real-world examples, rather than rough presentation. Repetition also helps remembering the key concepts.

The lecture must start and end on time. The rate of presentation should be no more than 10 words per 5 seconds. In addition, the amount of new material must be reasonable. ”Coverage is the enemy of understanding”. That is, it is very important to cover enough amount of material with enough details and reasonable pace of speaking.

Furthermore, the advantages of having student-centered classes outweigh the advantages of classes in which the teacher is the only speaker. Asking questions before expressing a new topic to challenge students, group discussions, practical exercises, lab experiments (if it is possible), and student’s teaching are examples of how to have student-centered classes.

Grading:

In grading it is very important to include different activities such as class attendance, class participation, assignments, projects, and exams rather. In addition, being flexible in grading and trying to help students in getting better grades, by considering student’s improvement within the semester or eliminating the worst grade before taking the average is preferred.

To analyze how to deal with a situation where an active student got a bad grade, we had a group discussion. In this discussion a teacher whose grading only relies on midterm and final tests had a student who was very active in class. However, the student got a bad grade in midterm exam. His scholarship depended on his grades. The question was what the teacher should do in this situation. Every of us had  different opinion, e.g., retaking the exam and giving some extra credit projects. Nevertheless, we all agreed that it is important to be fair to all students and also it is essential to change the syllabus at least for future courses as grading should not only depend on two tests’ results.

In general, this workshop was worth attending. However, it was not well-organized. For instance, the microphone did not work properly and most of the discussion was unheard by people, especially those who were sitting back. Moreover, the lecturer were not well-prepared for the presentation. They had to skip some slides as they ran out of time. Finally, the person who was managing the last discussion was not suitable for this purpose, as she did not seem to know the answer to problem (dealing with student’s grade).

Date: 10/30/2009

Location: Room 501 IS Building

Representing spatial information on map is a challenging task and requires knowledge on human cognition of their environment. Cartographers and map developers need to know which information must be presented on and which ones must be deleted from the map. Map schematization is a technique used for this purpose. Schematization is defined as the process of simplifying the representation of spatial information on a map. The objective of schematization is to represent only useful information to user in order to decrease the level of ambiguity. Schematization is application-dependent. For example, to schematize a map showing subway of a city, keeping information such as topology, direction, and distance are important, whereas having the information on exact course of tracks is not required. Techniques for schematizing maps require realizing the important features based on the application and ignoring the unimportant features.

The focus of the schematization discussed in this talk was mainly navigation application. However, navigation is a broad application and could be applied to different users and different environments (indoors or outdoors). For each user and environment the important information is different. For example, while displaying all buildings and landmarks on a map for drivers is distractive, it can be beneficial to pedestrians to navigate.

Date: 10/22/2009

Location: 4405 Gates & Hillman Centers

There are several acoustic models for speech recognition aiming at recognizing sound from noise. However, the problem with these models is their poor recognition performance. To obtain the performance similar to the human auditory system performance, which is very effective, some strict conditions are required. Current recognition techniques such as “adaptation”, “feature normalization”, “model compensated”, “feature enhancement”, “missing feature techniques”, and “matched condition training” have quite high performance in stationary noise. Nevertheless, these techniques do not take into consideration the nature of the noise while recognizing.  In this talk source separation algorithms and models based on non-negative matrix factorization and matrix deconvolution were presented. The methods were evaluated using both supervised and unsupervised learning.

Despite the advantages of this technique, experiments show that it cannot perfectly distinguish between different sounds such as between different genders or music and lyrics.

Date: 10/16/2009

Location:404 IS

Information delivery in wireless sensor networks is a challenging task as data may get lost during transmissions. Several parameters impact the performance of data transmission, including packet success ratio, time delay, and energy consumption. In order to improve the performance of data delivery, different approaches have been presented. Cost-based optimization is of such approach which aims at decreasing the amount of collisions that occur when a node communicate with other node that is already communicating with another one. In this approach, a cost model which is data transmission algebra schedule associated with combined time/energy costs is used.

A novel approach called GoLite, which is data delivery via light weight intelligent network was presented. The strategy used in GoLite is based upon the relationship between data rate and data loss, i.e., the less the rate of delivered data, the less the data loss. GoLite can be applied to wide variety of applications such as fire evacuation where the objective is rescuing as more people as possible while fire is spreading within a building. To demonstrate GoLite’s performance, it was evaluated and the result showed that it performs well.

Date: 10/08/2009

Location: 5317 Sennott Square

In this talk the challenges of using different techniques for argument formalism in law domain and a short demonstration were presented. An argument consists of a conclusion statement that is supported by a set/structure of premises. One possible approach for argument formalism is using knowledge representation. Knowledge representation is “a mechanism that allows drawing of inferences about the validity of a statement in a certain domain”, and is consisted of several statements about a domain. Using classical knowledge representation technique for argument formalism has some limitations. For instance, having the expressive knowledge representation leads to complex computations in that domain. To overcome its shortcomings, another approach, called Carneades, was presented. Carneades refers to an argument mapping application that provides different argumentation tasks (CAS, 2009). The advantage of using Carneades as an argument formalism are that (1) it has a sophisticated burden of proof model with which it resolves conflicts among competing arguments. Therefore, it enables the computation of inferences about the validity of a given statement; (2) it interfaces to a state-of-the-art burden of proof model as well as to other information, interchanges standards, and provides a set of readily available tools to develop a knowledge base.

This talk was rather technical and is quite hard for me to critique. However, in my opinion, the weakness of the talk was that it only discussed the limitations and advantages of the presented techniques without providing the background of each technique. In addition, in the Carneades demonstration not enough discussion on the relationship among the arguments were provided.

Reference

Carneades Argumentation System (CAS). Available at http://carneades.berlios.de/, [Accessed date 10/13/2009].

Date: 10/02/2009

Location: SENSQ 5317

In this talk a summary of Google approaches on searching products for sale was presented.  In order to provide users with searching service, Google gathers information on different products and save them in databases. The sources of information Google uses include merchants, web sites, users’ behaviors with respect to search interests, and text mining.  Obtaining information from each source has advantages and disadvantages. For instance, merchants, who are one of the sources of information, usually do not provide enough details, and they may not disclose real information on the product’s price. Web pages that are used for sources of information include merchants’ websites, product reviewers, etc. In order to extract information from web sites there are several techniques, such as annotating page titles by human and annotating DOM trees. In annotating DOM trees technique, information on web pages of a website is displayed in a tree, and then these trees are aligned together through a dynamic programming. In this technique the wrapper induction is very accurate and it can be deployed successfully in real world. However, due to the changes of the web sites’ contents and due to its high cost in terms of human efforts it is not considered as a perfect approach. Another approach is extracting information from text by classifying data fields using a machine learning classifier technique. This approach has pros and cons of machine learning techniques.