SEDS 539
Deep Learning
This course covers methods for designing and training deep neural networks. The course content includes the historical evolution of deep neural networks, their fundamental working principles and image classification and object detection and recognition in images using convolutional neural networks.
The fundamental aim of this course is for students to gain basic knowledge on deep neural networks and deep learning. By teaching the important aspects of convolutional neural networks, specifically, for the image classification task and making them into project assignments, the given theoretical knowledge is aimed to be applied in practice. The course’s target audience is graduate students with a background in machine learning.
Reference book(s):
• Michael A. Nielsen, “Neural Networks and Deep Learning”, Determination Press, 2015
• Ian Goodfellow, Yoshua Bengio and Aaron Courville, “Deep Learning”, MIT Press, 2016
Course Objectives:
To teach students the basics of deep neural networks and enable them to apply deep learning solutions on real problems.
| Week | Topics |
|---|---|
| 1 | History of artificial neural networks and introduction |
| 2-3 | Image classification with linear methods |
| 4 | Backpropagation |
| 5 | Training artificial neural networks I: data preprocessing, weight initialization and regularization, batch normalization and loss functions |
| 6 | Training artificial neural networks II: gradient checking, babysitting the training process, update methods, hyper-parameter optimization |
| 7-8 | Convolutional neural networks |
| 9 | Spatial localization and object detection with convolutional neural networks |
| 10 | Visualization and understanding of convolutional neural networks |
| 11 | Recurrent neural networks |
| 12 | Training convolutional networks in practice: data augmentation, transfer learning |
| 13 | Deep neural network coding frameworks |
Grading:
Final Exam %40
Midterm Exam %30
Practice Assignments %30
Course Learning Outcomes:
CO1 Ability to comment about historical evolution of biometric recognition
CO2 Ability to list the basic working principles of artificial neural networks
CO3 Ability to apply the learned fundamental principles to train artificial neural networks
CO4 Ability to classify images using convolutional neural networks
CO5 Ability to apply the acquired knowledge in practice using artificial neural network coding frameworks such as torch, TensorFlow.
Contribution of Program Learning Outcomes:
| PO1 | PO2 | PO3 | PO4 | PO5 | PO6 | PO7 | |
| CO1 | 1 | ||||||
| CO2 | 1 | ||||||
| CO3 | 1 | 1 | 1 | ||||
| CO4 | 1 | 1 | 1 | ||||
| CO5 | 1 | 1 | 1 | 1 |
Justification of the course:
It is an elective course of the Software Engineering and Data Science Master of Science Program and it is one of the important data science electives. The course introduces the fundamentals of deep learning by giving insight on how the deep neural networks are trained. The case study on image classification gives means to understand the mechanisms of training algorithms and the impact of learning parameters on the classification accuracies. With practical exercises, the students obtain hands-on experience with most popular deep learning libraries.
Overlapping with or complementing topics in courses:
This course requires a background knowledge in machine learning. So, its topics do not overlap with SEDS 537 Machine Learning but builds on top of it.
