| |
Apr 23, 2024
|
|
|
|
|
CS 6890 - Deep Learning
This course introduces algorithms for the unsupervised learning of feature representations in the context of deep architectures. Basic features, such as edge detectors in computer vision, will be induced automatically from unlabeled data using sparse autoencoders and then assembled into increasingly more complex feature representations by greedy layering in deep learning architectures. Major topics include Feedforward Neural Networks, Backpropagation, Sparse Autoencoders, Code Vectorization, PCA and Whitening, Self-Taught Learning, Deep Networks, Linear Decoders, Convolution and Pooling, Sparse Coding, Independent Component Analysis, and Canonical Correlation Analysis.
Requisites: CS 6830 or permission
Credit Hours: 3
Repeat/Retake Information: May not be retaken.
Lecture/Lab Hours: 3.0 lecture
Grades: Eligible Grades: A-F,WP,WF,WN,FN,AU,I
Learning Outcomes:
- Student will be able to demonstrate the utility of UFLDL in the context of a concrete research project.
- Students will be able to derive Principal Component Analysis and express it as the solution of a linear auto-encoder.
- Students will be able to derive the gradients of a number of standard objective functions in machine learning, such as sum-of-square errors or cross-entropy.
- Students will be able to describe major approaches for learning feature representations.
- Students will be able to distinguish among supervised, semi-supervised, and self-taught learning.
- Students will be able to evaluate the impact of unsupervised feature learning on the performance of supervise machine learning models.
- Students will be able to explain the advantages of deep learning architectures over shallow architectures.
- Students will be able to implement and evaluate sparse and denoising autoencoders.
- Students will be able to implement and explain the utility of convolution and pooling.
- Students will be able to implement linear decoders.
- Students will be able to list the difficulties one may encounter when training deep architectures and outline solutions to alleviate them.
- Students will be able to read relevant publications on their own in order to summarize and evaluate existing unsupervised feature learning and deep learning (UFLDL) techniques.
Add to Portfolio (opens a new window)
|
|
