Introduction to Neural Networks
Course facts
- Payment: Free, no certificate
- Credits: Non Credit
- Learning style: Self-paced, start anytime
- Resources: Lecture Notes; Problem Sets
MIT OpenCourseWare details
- As taught in: Spring 2005
- Department: Brain and Cognitive Sciences, Physics
| Course Details | ||||||
|---|---|---|---|---|---|---|
|
||||||
| Outcomes | ||||||
|
||||||
| Price & Terms | ||||||
|
||||||
| Technical | ||||||
|
||||||
| Specification | ||||||
|
||||||
Neural computation from a neuroscience view: perceptrons, recurrent networks, associative memory, backpropagation and reinforcement learning.
How can networks of neurons compute and learn? This class uses mathematics to study feedback and nonlinearity, focusing on recurrent networks because feedback loops dominate the brain’s connectivity. It covers perceptrons, linear and antisymmetric networks and associative memory. It then treats PCA, backpropagation, stochastic gradient descent and reinforcement learning, with ten problem sets that use real data.
What You Will Achieve
- Model neurons using integrate-and-fire and rate equations
- Analyze recurrent networks with Lyapunov functions
- Implement perceptrons and backpropagation
- Compare Hebbian, delta rule and reinforcement learning
Program Overview
Perceptrons and Linear Networks
The first lectures build the units.
- From spikes to rates
- Perceptrons: simple and multilayer
- Lateral inhibition and feature selectivity
Recurrent Dynamics
The middle lectures study feedback.
- Objectives and optimization
- Antisymmetric networks
- Associative memory
- Multistability
Learning
The later lectures address how networks change.
- PCA
- Backpropagation
- Stochastic gradient descent
- Reinforcement learning
Course Completion
Job Outlook
This grounding helps researchers in computational neuroscience and those who want a biological perspective on machine learning.
Good for: Students with linear algebra, multivariate calculus, probability and programming who are interested in brains and computation.
Keep in mind: It is not a guide to current deep learning tools. Problem sets assume MATLAB, and solutions are not listed.
Bottom line: A mathematical look at neural computation, notable for its emphasis on recurrent dynamics.
Frequently Asked Questions
Which book is recommended?
Introduction to the Theory of Neural Computation by John Hertz, Anders Krogh and Richard Palmer, published in 1991.
What do the problem sets use?
Topics range through integrate-and-fire neurons and Lyapunov functions, and later sets supply data including faces and an abridged MNIST set in MATLAB format.
What does it leave out?
The biophysics of single neurons, taught in a separate class, and much statistical pattern recognition, as the syllabus points to other courses for these.
User Reviews
Write your review
You must be logged in to post a review.
Reviews can only be posted by registered learners. Log in or create a free account, then you will return here to post.
There are no reviews yet.