CS 559: Deep Learning

Fall 2023

• Instructor: Hamdi Dibeklioglu
• Office hours: by appointment / online
• Class hours:
• Wednesday 08:30-10:20, Friday 13:30-15:20
• While every other week we will have 2+2 hours of lectures, other weeks there will be one 2-hours lecture. Please see the schedule below for the dates of the lectures.
• Class room:
• EB-104

UPDATE:

• All important dates (including homework, project/project progress, and survey submission/presentation deadlines, and midterm date) have been announced.
• Homework details have been announced, please see the homework section below (07.11.2022).
• When you submit your proposals for survey and project, please make sure that you use the required keyword in the subject line of the corresponding email. For the subject line keywords please see the Important Dates section below.
• Lectures start as of 15 September 2023.

Catalog Description: Overview of machine learning and its applications. Loss functions, numerical optimization and back-propagation. Fundamentals of feedforward neural networks. Modern architectures and techniques for training deep networks. Convolutional neural networks: basics, visualization, and techniques for efficient spatial localization in images. Recurrent neural networks and their variants. Applications of recurrent neural networks in language and image understanding, and image captioning. Recent advances in generative models learning, generative adversarial networks and variational auto encoders. Unsupervised and self-supervised representation learning. Deep reinforcement learning.

Recommended Books:

• I. Goodfellow, Y. Bengio, A. Courville. Deep Learning, MIT Press, 2016. [Available Online]
• K. P. Murphy. Machine Learning: A Probabilistic Perspective, MIT Press, 2012.
• C. M. Bishop. Pattern Recognition and Machine Learning, Springer, 2006.

Assessment Methods:

Any of the following will directly result in an F grade:

• not submitting a project or homework (including report),
• not preparing/presenting a survey on the pre-scheduled date,
• being absent in the midterm,
• being absent in a project presentation.

Passing Grade: No predetermined grade to pass the course.

Makeup Policy: Medical report holders will be entitled for the midterm make up. Makeup exam will be comprehensive.

Homework:
Please follow this link to access the homework details. Homework details/materials are accessible only within Bilkent network. Use VPN to access from home.

Literature Survey and Presentation:

• Groups of two students will choose a topic related to deep learning, and prepare a short survey on it.
• Surveys should be based on about 10 papers (report: 5 pages max.).
• You will make a presentation on your survey in class. The presentation should be in parallel with your report.
• Survey topics should be confirmed first. Very similar topics to others’ will not be allowed (priority: first come, first served).
• Your chosen survey topic and a few lines explanation (indicating group members) should be sent to dibeklioglu@cs.bilkent.edu.tr by 29 September 2023, 23.59 (Turkey time).

Project:

• By groups of two or three students.
• Explore novel applications of contemporary deep learning techniques or develop novel deep learning techniques.
• Projects related to your research topics are encouraged.
• Three stages:
• Proposal: one-page description of the project topic and the planning for the project (indicating group members) should be sent to dibeklioglu@cs.bilkent.edu.tr by 2 October 2023 (23:59 Turkey time)
• Progress report & presentation (report: 2 pages max.)
• Final report & presentation (report: 5 pages max.)
• You are allowed and encouraged to use mainstream deep learning libraries like TensorFlow, PyTorch, Torch, etc.

Template for the Reports:
All reports (including homework report) must be prepared using the IEEE double-column transactions article template (i.e. "bare_jrnl.tex").

Important Dates:

Tentative Schedule & Lecture Notes:
Lecture notes below are downloadable only within Bilkent network. Use VPN to access from home.

Week	Topic	Dates	Lecture Notes
1	Introduction	15 September 2023 (13:30-15:20)
2	Basics, Loss Functions	20 September 2023 (08:30-10:20) 22 September 2023 (13:30-15:20)
3	Optimization Feedforward networks and training (1)	27 September 2023 (08:30-10:20) 29 September 2023 (13:30-15:20)
4	Feedforward networks and training (2)	04 October 2023 (08:30-10:20)
5	No Lecture	-
6	Convolutional neural networks Spatial localization and detection	18 October 2023 (08:30-10:20) 20 October 2023 (13:30-15:20)
7	Segmentation Understanding and Visualizing CNNs	25 October 2023 (08:30-10:20) 27 October 2023 (13:30-15:20)
8	Recurrent Neural networks	01 November 2023 (08:30-10:20)
9	Word Embeddings and Language Models	08 November 2023 (08:30-10:20)
10	Word Embeddings and Language Models, Unsupervised Learning and Generative Models	15 November 2023 (08:30-10:20)	>>
11	Project Progress Presentations	22 November 2023 (08:30-10:20)
12	Unsupervised Learning and Generative Models Midterm	29 November 2023 (08:30-10:20) 01 December 2023 (13:30-15:20)	>>
13	Literature Survey Presentations	06 December 2023 (08:30-10:20) 08 December 2023 (13:30-15:20)
14	Deep reinforcement learning Project Presentations	13 December 2023 (08:30-10:20) 15 December 2023 (13:30-15:20)
15	Project Presentations	20 December 2023 (08:30-10:20)