LECTURE 71 WHAT IS MACHINELEARNING
We provide a large number of examples from the sort of thing we
want the computer to learn Instead of providing the computer a set
of instructions step bystep the computer figurs out its own program
examples sortinghandwritten adresses computer chess self driving cars
A ML problem is suitable when
we can't solve it explicitly
approximate solutions are fine
limited reliability predictability interpretability is fine
plenty ofexamples to learnfrom
ML provides systems the ability to automatically learn and improve
from experience without beingexplicitlyprogrammed
Most of this Ml course is about offline learning The main idea is
that we seperate learningpredicting and acting
We start with a dataset of examples instances Then feed these
instances into a learningalgorithm that produces a model
Test the model to see if it works by checking its predictions
We want Solutions that are applicable across domains not one
solution for chess one for cars etc Hence we abstract the
learning task to an abstract task e g Classification regression
and then develop an algorithm for these abstract tasks
,Two types of abstract tasks
Supervised abstract tasks
provide examples of input and output
learn to predict the output from an unseen input
Classification
regression
unsupervised abstract tasks
provide input only
find any pattern that explains something about the data
clustering
density estimation
generative modeling
Basic recipe for offline ML
1 Abstract the problem to a Standard task
2 Choose instances and their features
3 Choose model class
4 Search for a good model
,LECTURE 12 CLASSIFICATION
Classification is the task of assigning a class to an example
That is one of a finite number of categories
The basic framework of classification has data that
consists of examples instances of things we are trying to
learn something about Per instance we measure some
feature numeric categorie Finally we have the target value
the thing we are
trying to learn In classification this is
always a Class of possible values
3 examples of classification algorithms
linear classifier
With 2 features above the line is
ei predicted and below the line
Decision tree classifier
42 yo Per target value 2 classes are possible
la Gb If the first eg is yes and the
Yesh yo gesp y no
a b a b second not the class b is assigned
K nearest neighbours KNN
The algorithm searches for the k in the example
_jij KU nearest neighbours and assign the class
that is most frequent in that set
, LECTURE 1.3 OTHER ABSTRACT TASKS REGRESSION
CLUSTERING DENSITY ESTIMATION
Regression supervised
Works exactly the same as Classification except we're
predicting a number instead of a Class
The features of a particular instance i are denoted by the
vector xi The corresponding true label given the data is ti
The models prediction is given by f xi
Goal is to get f xi as close as possible to ti
To determine how good a model is
we use a lossfunction
ii Igression Mean squared error Insel loss
loss If IF
We can also use the decision tree principle to perform
a regression tree Just like classification
regression giving us
we can even apply KNN to regression as well
Clustering unsupervised
Goal here is to split the instances into a number of
clusters This number is usually given by the user in advance
One example of clustering is k means start
bij choosing
k random points in the feature space the means Assign
each point to the closest mean giving k clusters Next re compute
the means for each cluster and re assign Points to its closest
mean Repeat this process until the means stop moving
We provide a large number of examples from the sort of thing we
want the computer to learn Instead of providing the computer a set
of instructions step bystep the computer figurs out its own program
examples sortinghandwritten adresses computer chess self driving cars
A ML problem is suitable when
we can't solve it explicitly
approximate solutions are fine
limited reliability predictability interpretability is fine
plenty ofexamples to learnfrom
ML provides systems the ability to automatically learn and improve
from experience without beingexplicitlyprogrammed
Most of this Ml course is about offline learning The main idea is
that we seperate learningpredicting and acting
We start with a dataset of examples instances Then feed these
instances into a learningalgorithm that produces a model
Test the model to see if it works by checking its predictions
We want Solutions that are applicable across domains not one
solution for chess one for cars etc Hence we abstract the
learning task to an abstract task e g Classification regression
and then develop an algorithm for these abstract tasks
,Two types of abstract tasks
Supervised abstract tasks
provide examples of input and output
learn to predict the output from an unseen input
Classification
regression
unsupervised abstract tasks
provide input only
find any pattern that explains something about the data
clustering
density estimation
generative modeling
Basic recipe for offline ML
1 Abstract the problem to a Standard task
2 Choose instances and their features
3 Choose model class
4 Search for a good model
,LECTURE 12 CLASSIFICATION
Classification is the task of assigning a class to an example
That is one of a finite number of categories
The basic framework of classification has data that
consists of examples instances of things we are trying to
learn something about Per instance we measure some
feature numeric categorie Finally we have the target value
the thing we are
trying to learn In classification this is
always a Class of possible values
3 examples of classification algorithms
linear classifier
With 2 features above the line is
ei predicted and below the line
Decision tree classifier
42 yo Per target value 2 classes are possible
la Gb If the first eg is yes and the
Yesh yo gesp y no
a b a b second not the class b is assigned
K nearest neighbours KNN
The algorithm searches for the k in the example
_jij KU nearest neighbours and assign the class
that is most frequent in that set
, LECTURE 1.3 OTHER ABSTRACT TASKS REGRESSION
CLUSTERING DENSITY ESTIMATION
Regression supervised
Works exactly the same as Classification except we're
predicting a number instead of a Class
The features of a particular instance i are denoted by the
vector xi The corresponding true label given the data is ti
The models prediction is given by f xi
Goal is to get f xi as close as possible to ti
To determine how good a model is
we use a lossfunction
ii Igression Mean squared error Insel loss
loss If IF
We can also use the decision tree principle to perform
a regression tree Just like classification
regression giving us
we can even apply KNN to regression as well
Clustering unsupervised
Goal here is to split the instances into a number of
clusters This number is usually given by the user in advance
One example of clustering is k means start
bij choosing
k random points in the feature space the means Assign
each point to the closest mean giving k clusters Next re compute
the means for each cluster and re assign Points to its closest
mean Repeat this process until the means stop moving
