Week 6

Página 1

when errors exist in ML algorithm ?

- add more training data

- add more features

- remove features

- try polynomial fit

- try increasing lambda

-Try decreasing lambda

Use diagnostic: that tells how accurately the ML algorithm fit/don't fit the data;

Página 2

hypothesis testing:

divide data into training & test data

- Use algorithm to learn parameter Theta from the training set;

-compute test set error

- 1/2m sigma (htheta -y)^2 (for linear regression)

- for logistic regression:

-use misclassification error :

proportion of test data that was misclassified

Página 3

Model Selection :

Choose models with different degrees from linear regression (d=1) to polynomial (d=10);

Compute the learning parameter theta from Training set

Compute the error from the Test set;

error from Test set is important because : at higher degrees , polynomial might overfit the current data on the training sets

model selection:

divide data into training, cross validation and test sets

run models on training;

verify error on cross validation; pick a specific model

test the specific model on test data

Página 4

bias and variance:

compute the training error and cross validation for models with different degrees of polynomials;

Jtrain

Jcross ;

if jtrain and jcost are both high , High bias issue - Underfitting

jtrain is low and jcost is high, high variance issue - Overfitting

Página 5

choosing learning parameter Lambda:

choosing the learning parameter Lambda:

determine different lambdas (0 , 0.01...10)

run thru the different models with different degrees

for different lambdas , go thru different models to learn thetas

Compute the cross validation error using the learned Θ (computed with λ) on the JCV(Θ) without regularization or λ = 0

Select the best combo that produces the lowest error on the cross validation set.

Using the best combo Θ and λ, apply it on Jtest(Θ) to see if it has a good generalization of the problem

Página 6

Learning Curve:

plot error as function of m;

ex: fitting quadratic function to m=1 ,2 ,3 data sets will result in low training error

for m > 3 . training error starts to increase

for cross validation set , error is high with m -1,2,3 ; and error starts to decrease for m >3

another example : fitting linear regression : high bias case

fit the best possible straight line;

low training set : low training error; high cross validation error

cross validation error plateaus soon with high error value;

with more data : training error increases to high value and value closer to cross validation error

Experiencing high bias:

low training set : low training error; high cross validation error

high training set : high training error ; high cross validation error

If a learning algorithm is suffering from high bias, getting more training data will not (by itself) help much.

Experiencing high variance:

high order polynomial fit (>10)

Low training set size: Jtrain(Θ) will be low and JCV(Θ) will be high.

Large training set size: Jtrain(Θ) increases with training set size and JCV(Θ) continues to decrease without leveling off. Also, Jtrain(Θ) < JCV(Θ) but the difference between them remains significant.

If a learning algorithm is suffering from high variance, getting more training data is likely to help.

Página 7

neural networks and fitting:

small nn : prone to underfitting

large nn : prone to overfitting ; use lambda to address it

Página 8

machine learning system design:

building spam classifier:

supervised learning : y=1 => spam ; y=0 => not spam;

create a vector list of words : spam : non-spam words : X features: buy, discount, Andrew, etc

assign : Xj =1 => if word appears ; Xj =0 : if not appears;

options to build better classifier:

Collect lots of data (for example "honeypot" project but doesn't always work)

Develop sophisticated features (for example: using email header data in spam emails)

Develop algorithms to process your input in different ways (recognizing misspellings in spam).

Página 9

machine learning system design:

error analysis:

when starting on new mc learning problem:

1) build quick and dirty solution

2) test it on cross validation set :

3) plot learning curves to test if more data , more features are likely to help;

3) manually analyze errors of cross-validation set to spot where errors were made

we should try new things, get a numerical value for our error rate, and based on our result decide whether we want to keep the new feature or not.

Página 10

precion and recall:

precision = true +/(true + plus false +)

recall = true +/ (true + plus false negative)

metric for threshold :

F score = 2PR/(P+R)

for different threshold : find one that has high f score;