# Extra Flashcards

1
Q

Why do we want non-polynomial functions

A

because polynomial functions have multiple low or high points on the graph = multiple answers.

2
Q

If we have an error of 500 in regression how do we correct that?

A

Gradient descent will use backpropergation to adjust the theta values to minimize the output by 500

3
Q

Why do we square cost functions in regression

A

we want to avoid some data in the sum function to cancel out other values as they might be negative. (when adding to the sum function)

4
Q

Is the XOR function linear or non-linear?

A

It is of course, non-linear it.

It returns a binary result, 1 or 0, dependent on whether

5
Q

What is key and value in a dictionary in terms of Markov Chains?

A

In Markov Chains the key is the word you want to generate the following word for, and the value is a list of possible words that can follow.

6
Q

What is a decision tree?

A

It is a flow-chart like model, where each node is a test on an attribute. e.g. left is yes and right is no.

It is used to make decisions.

Supervised learning as you deal with decisions that are predetermined and you know the cases.

7
Q

What is the random forest algorithm

A

It is a classification and regression algorithm for learning. It is a supervised learning method.

8
Q

What is K-nearest neighbour?

A

it is a classification algorithm that assumes that values with similar values are similar

9
Q

When a Deep Neural Network model learn, what is it essentially the method of how it learns?

A

It minimizes which weight and biases minimize a certain cost function (sum of error)

10
Q

What is backpropagation

A

Backpropagation is how we adjust our weights after we have identified how much error we have with the cost functions.

11
Q

If your model is 100 % exact on your training set what does that imply?

A

It implies that something is wrong and that you need to look into why

12
Q

What can you do with your model if your model is predicting too precisely to the training set?
Think in terms of nodes.

A

You can remove some layers, causing the DNN to have less of a chance of getting the exact weights.

13
Q

What can you do if you want to make it harder for your model to predict too closely to the training set

A

You can add regularization which is basically taking the sum of the weights and dividing that with the cost function. Then whenever the model changes the weights to get a better prediction the change in the weights are not accounted for making it less precise.

14
Q

Why dont you want your model to predict 100 % precise

A

You dont want it to not be able to predict outside of the training set. so outside with new values. It can have learned the errors as features also