Curriculum
  /  
Descriptions
MSAI 395, 495: Deep Learning Foundations From Scratch

Quarter Offered

None ; TBD

Description

The​ ​last​ ​decade​ ​has​ ​seen​ ​a​ ​resurgence​ ​of​ ​deep​ ​neural​ ​networks​ ​as​ ​the​ ​machine​ ​learning  tool​ ​of​ ​choice​ ​for​ ​many​ ​application​ ​areas​ ​including​ ​computer​ ​vision,​ ​reinforcement​ ​learning,​ ​and​ ​natural  language​ ​processing.​ ​This​ ​course​ ​offers​ ​a​ ​holistic​ ​and​ ​hands-on​ ​introduction​ ​to​ ​deep​ ​networks,​ ​their​ ​many  varieties​ ​and​ ​applications,​ ​as​ ​well​ ​as​ ​the​ ​algorithms​ ​used​ ​to​ ​train​ ​them.​ ​In​ ​this​ ​course​ ​we​ ​discuss​ ​the  foundations​ ​of​ ​Feedforward​ ​Networks,​ ​Convolutional​ ​Networks,​ ​and​ ​Recurrent​ ​Networks,​ ​as​ ​well​ ​as​ ​their  usage​ ​within​ ​popular​ ​Reinforcement​ ​Learning​ ​frameworks.​ ​​Using​ ​real​ ​datasets​ ​and​ ​basic​ ​Python​ ​libraries​ ​for  data​ ​manipulation,​ ​vector/matrix​ ​algebra,​ ​and​ ​automatic​ ​differentiation​ ​students​ ​will​ ​code​ ​up​ ​-​ ​from​ ​scratch​ ​-  popular​ ​examples​ ​of​ ​each​ ​such​ ​network​ ​architecture​ ​and​ ​apply​ ​them​ ​to​ ​applications​ ​in​ ​natural​ ​language  processing,​ ​computer​ ​vision,​ ​speech​ ​recognition,​ ​and​ ​automatic​ ​control.

  • This course fulfills the Technical Elective requirement.

INSTRUCTORSDr. Reza Borhani and Dr. Jeremy Watt

COURSE OUTLINE:

1. Deep Feedforward Networks

  • a. Supervised learning recap
  • b. Feedforward networks in the context of function approximation
  • c. Simple recipes for building deep networks
  • d. Deep nets vs Kernel methods
  • e. How deep networks came back

2. Technical Issues with Deep Networks

  • a. Mathematical optimization recap
  • b. Automatic differentiation, implementations in Python
  • c. First order stochastic gradient methods
  • d. Regularization techniques
  • e. Technical tricks

3. Convolutional Networks

  • a. From fixed feature extractors to convolutional networks
  • b. Composing filters and functions
  • c. Applications in computer vision and speech processing
  • d. Transfer Learning

4. Reinforcement Learning

  • a. Reinforcement Learning fundamentals
  • b. Basic Q-Learning, enhancements, and generalizability
  • c. Injecting function approximators into the Q-Learning scheme
  • d. Policy gradient method
  • e. Applications to automatic control and Atari game AI

5. Recurrent Networks

  • a. Recurrence relations
  • b. Parameterized recurrence relations and feedforward networks
  • c. Deriving vanilla recurrent networks
  • d. Popular architectures, Long-Short-Term-Memory
  • e. Applications in NLP and speech recognition

6. Deep Representation Learning

  • a. Unsupervised learning recap
  • b. Nonlinear extensions of PCA, autoencoders
  • c. Unsupervised generative models

COURSE HAND-OUTS: This course has no required textbook. Handouts authored by the instructors will be made freely available to students for notes. In addition a small number of relevant sections from the author’s textbook Machine Learning Refined (Cambridge University Press) will be made freely available to students.

PROBLEM SETS: 6-7 problem sets will be assigned and graded.

COURSE PROJECT: There will be one course project worth 25% of the final grade.

COURSE GRADE: Final grades for the course are based on homework (75%) and course project (25%)