This topics course aims to present the mathematical, statistical and computational challenges of building stable representations for high-dimensional data, such as images, text and data. We will delve into selected topics of Deep Learning, discussing recent models from both supervised and unsupervised learning. Special emphasis will be on convolutional architectures, invariance learning, unsupervised learning and non-convex optimization.
Course Outline (tentative)
1st part: Convolutional Neural Networks
- Invariance, stability.
- Variability models (deformation model, stochastic model).
- Scattering networks
- Group Formalism
- Supervised Learning: classification.
- Properties of CNN representations: invertibility, stability, invariance.
- covariance/invariance: capsules and related models.
- Connections with other models: dictionary learning, LISTA.
- Other tasks: localization, regression.
- Embeddings (DrLim), inverse problems
- Extensions to non-euclidean domains
- Dynamical systems: RNNs.
- Guest Lecture
2nd part: Deep Unsupervised Learning
- Autoencoders (standard, denoising, contractive, etc etc)
- Variational Autoencoders
- Adversarial Generative Networks
- Maximum Entropy Distributions
- Guest Lecture
3rd part: Miscellaneous Topics
- Non-convex optimization for deep networks
- Stochastic Optimization
- Attention and Memory Models
- Open Problems
Our main resource will be a github course project
which will contain updated references, pointers to papers and lecture slides.
Linear Algebra, Analysis, Probability, some notions of Signal Processing, and Numerical Optimization.
This course grading will have two components:
- Paper reviewing (30%): you will be assigned two papers each, and you will be asked to produce a review following the standards of journal/conference publications.
- Final Project (70%): It can consist in either of these three options:
- Oral presentation of a recent paper to the class,
- Tiny research project,
- Contribute to an open-source software package (Torch, Caffe or Theano).
Final Project Proposals are due by email (email@example.com) on April, 1st. Final projects are individual, unless there is a compelling reason for teaming up.
Tuesdays from 4pm to 6pm, Evans 419, or by appointment.
Students are expected and encouraged to collaborate and share coursework. However, assignments and final projects should be conducted individually, unless there is a compelling reason to collaborate (that I should approve previously).
The syllabus page shows a table-oriented view of the course schedule, and the basics of course grading. You can add any other comments, notes, or thoughts you have about the course structure, course policies or anything else.
To add some comments, click the "Edit" link at the top.