CSE 591 Introduction to Deep Learning in Visual Computing Instructors: Baoxin Li & Ragav Venkatesan Computer Science & Engineering Course notes will be made available online January, 2017 1
Lecture Overview A talk given last year: deep learning, vision & others. Course information, syllabus and logistics. Basics in image representation. 2
About This Course This is a new course we just started to offer this spring. We (both some of our faculty at ASU and my group) have many research activities related to deep learning. Ragav and I just completed a concise book on the same topic. Ragav will be a co-instructor (and will deliver 7 lectures); I will also be present in all lectures and ask questions (unless when traveling) To make this an interactive seminar course. My Office: BY554/502. Office hours: Monday 9am-10am @ BY554, Wednesday 11am-noon @BY502, or by appointment. Other temporary cancellations due to travel or other urgent university business will be announced on Blackboard and made up. Co-instructor/TA: Ragav Venkatesan Email: ragav.venkatesan@asu.edu Office: TBD Office hours: M/W 1-2pm 3
Many other students in my group have worked on research tasks related to this course. I will likely recruit them to enhance the delivery of this course They may participate in discussion, grading your presentation/project, etc. They may give short guest presentations. 4
Prerequisites You need to have a working knowledge of calculus, linear algebra and basic probability theory. Ideally, you should have taken at least one graduate-level machine learning classes like CSE569 or CSE575 at ASU. Proficiency in Python programming is needed for the course project and homework assignments. plus You will need some physical vigor for sitting here for 2.5 hours listening, presenting, & participating in discussion. 5
Course Information Grade Center on the Blackboard will be used for documenting the assessments, but most course materials will be hosted here: http://www.public.asu.edu/~rvenka10/cse591 Textbook: A Concise Guide to Modern Neural Visual Computing, (in press), CRC Press, by Ragav Venkatesan & Baoxin Li. The pre-print chapters of the book will be provided to students in this class. Other references Deep Learning, Ian Goodfellow and Yoshua Bengio and Aaron Courville. Neural Networks for Pattern Recognition, Christopher Bishop. Online book: http://neuralnetworksanddeeplearning.com/ Michael Nielson. A pool of recent research papers to be posted. 6
Course Information The mini-project assignments and the project will require programming work. You need to learn to do (if not knowing already) Python programming. Note: This is not a programming course, so we will not grade your code, but only its outcomes (including speed performance). Late submissions of the assignments will not be accepted. 7
Assessment Homework (tentatively 6 mini projects) 24%: Pop quizzes or attendance 6% We may either have quizzes or simply take attendance, for 4 times at random times. Each time accounts for 1.5%. If you are missing during a quiz or when we take the attendance, you lose that 1.5%. No make-up will be given, even if missing the class is due to, e.g., valid medical reason, since this part is for attendance tracking. Medical emergency/conditions may qualify you for special considerations like late withdrawal, incomplete grade, etc., but not automatically earn you attendance credit. Paper presentation 5% 3 or 4 students will form a group to present a paper. Peer-grading may be employed in this part. 8
Assessment Midterm exams: two of them, 12% each. Each midterm will cover only materials in preceding period. Final exam: 11%, on the official final exam day (will be assigned by the university). The final exam is supposed to be comprehensive, and it may contain problems testing your understanding the papers presented during the semester. All the exams (midterms & final) will be closed-book, but 1/2/3 respectively cheat sheet(s) will be allowed for midterm1/2/final exam. Solutions for exams may be discussed during the lectures but will not be posted. You will have a chance of looking at your graded exam paper but they will not be returned to you. 9
Assessment Project: 30% You may form a 2-person group for the project; You may do it alone but we will not give any extra credit or special grading consideration for that. No more than 2 will be allowed in a group. You may propose your own topic (but we need to approve it). Project topics will be finalized within 2 weeks of the first midterm (by Feb. 17 or so). We will provide some topics, if you have no topic of your own. You will need to submit a mid-term report (5%), summarizing your progress on the project by the due day of the report (due sometime late March or early April; detailed requirement on the report to be determined after we finalize the topics). 10
Grading Scheme The following cutoffs represent what will be likely used to generate the letter grade A, B, C and D: A: >= 85% B: >= 75% C: >= 60% D: >= 50% (Plus/Minus will be interpolated accordingly but A+ is rarely used except for truly outstanding cases.) No curve fitting. 11
Email Policy All email communications need to follow the guidelines enumerated below Email communication regarding this class MUST include in the subject line the prefix CSE 591: (For example, the subject line of your email may read CSE 591: Question on HW1). Every email must also cc Ragav (unless there is a specific and clear reason why he should not be cc'ed). (Note: Ragav is an official class staff member and has full access to the Grade Center on Blackboard.) Emails will be read once a day, M-F. I will request Ragav to answer all email he is copied on, unless he feels that my answer is needed. Email should be clear, self-contained, and to the point. Email should not ask questions whose answers are obviously shown in the course syllabus, classnotes/class materials, or other materials on Blackboard. 12
Email Policy (continued) Avoid asking questions in email that should be raised either in class, or in individual consultation with the TA/instructor during office hours. These include questions of an excessively conceptual nature, and questions that require an unreasonable amount of time from the instructor/ta to answer/explain via email. A good rule of thumb: if your question cannot be answered in a short paragraph, then it is not appropriate for email. Emails that do not follow these guidelines may not be replied by the TA/instructor. If your email goes unanswered more than one day after you sent it, check if you forgot following these guidelines. 13
Discussion Group https://groups.google.com/forum/#!forum/asu-spring2017- cse591-dl4cv For discussion, ask and answer question. Can be useful to address many common questions (e.g., clarification on assignments); more effective than emails. 14
Academic Integrity A perceived lack of academic integrity undermines a school s reputation, and devalues your degree. ASU Academic Integrity Policy: http://provost.asu.edu/academicintegrity All violations for which a penalty is assigned must be reported to the Dean s office. This is NOT a matter of faculty discretion, but a university-mandated legal requirement. All the assignments, quizzes/exams are individual work except stated otherwise; the project and paper presentation allow collaboration only within the same group. Additional info from Office of Graduate Education: http://graduate.asu.edu/beintheknow See a flier. 15
Some efforts taken to ensure Academic Integrity During exams, your seat will be assigned. You sit wherever we ask you to sit. We may use different versions of exam papers in the same exam. We will run your report/project code through plagiarism detection software. Take this seriously, as we did this before. You cannot easily fool those software by small tricks like simply change the variable names. 16
Topics To Be Covered & Tentative Schedule Introduction to visual representation & fundamentals of machine learning Neural networks & backpropagation Optimization techniques for neural networks General deep learning paradigms (CNN, auto-encoder, GANs etc.) Modern convolutional neural networks Software implementation of deep learning Selected recent advances in deep learning == A tentative class schedule is summarized in this table. 17
in addition The important part of paper presentation Some of the papers will be very recent, e.g., we may pick some from the AAAI conference to be held next month. Time permitting, we might plan for guest presentations. 18
Common Qs & As Will you give grades lower than B? Yes. Lower than 75% = lower than B. I have been a straight A student, but why it looks like I m only in the B range Good question. I got only a C, but I rely this course to satisfy my degree requirement. I cannot graduate with a C grade. Can I do an extra assignment to improve my grade? No. Emails asking such type of questions will not be responded. I m only 0.5% below the B cut-off, can you please change my grade from B- (B minus) to B? No. Emails asking such type of questions will not be responded. 19
Common Qs & As I got only a D, which put me on academic probation. Is there anything extra I may do to improve my grade, please? No. We go strictly with the published syllabus. Emails asking such type of questions will not be responded. I missed the exam. Can I have a make-up one? No, unless you have official documents supporting a genuine emergency. I have multiple assignments due this week and thus I couldn t finish the assignment. Can I get an extension to turn in this homework? No. Dealing with multiple assignments due at the same time is part of the study life. Please plan ahead and don t wait until the last minute and then ask such questions. 20
Additional questions or comments 21