Gradient descent

Question 1

Define gradient descent

Answer 1

Gradient descent is a strategy for finding the valleys of a function, if the amount of error in a model can be written as a function, gradient descent can be used on it to reduce the error until it reaches 0.

Question 2

How to use gradient descent on an error function

Answer 2

1. Find the derivative - this tells us the gradient of the function for every parameter value.
2. put a minus in front of it. - in order to follow the gradient down hill to reduce the error.
   The result is a maths rule which can be used to input parameter values to find out how to reduce the error.

Question 3

What real life problems are explained by gradient descent?

Answer 3

• Function minimisation (maths)
• Energy landscapes (physics)
• Fitness landscapes (evolution) where fitness in terms of evolutionary success is mapped against genotypes.
• Ontogenetic landscapes (psychology)
• Epigenetic landscape (biology)
• Error surface (neural networks)

Question 4

What explanations are derived from gradient descent?

Answer 4

homeostasis. 
Supervised learning (the rescorla- wagner model)