Queen's School of Computing

CISC 473/3.0 Deep Learning

Original Author: David Skillicorn
Last Revised: 2019-03-20

Calendar Description

Design of deep neural networks based on leading-edge algorithms such as Restricted Boltzmann Machines, Recurrent Neural Networks, Convolutional Neural Networks, Long-Short Term Machines. Autoencoding as a clustering technique. Applications to prediction problems in natural language and images.

Prerequisites: CISC 371/3.0

Learning hours: 120 (36L; 84P)

Degree Planning

  • This course is required for the Data Analytics focus of the COMP degree plan.

Topics

  • Introduction to deep networks, review of optimisation for learning, backpropagation (1 week)
  • Restricted Boltzmann machines (RBMs), stacking RBMs (1 week)
  • Autoencoding as a clustering technique, Stacked Denoising Autoencoders (2 weeks)
  • Vanishing gradient problem, extending backpropagation to other network structures (1 week)
  • Recurrent neural networks, applications to natural language prediction problems (2 weeks)
  • Convolutional neural networks, applications to natural language and image prediction problems (2 weeks)
  • Long-short term machines (LSTMs), biLSTMs, applications to natural language problems (2 weeks)
  • Matching techniques to problem domains, comparative performance (1 week)