Sei sulla pagina 1di 9

Sir USMAN GHANI LECTURES

(Lecture 1-Lecture 1.6)

Lecture 1
1. Introduction to Machine Learning

2. Our basic focus is on machine learning basic discussion and also we


discuss the advancement in machine learning.

3. Specifically we will be focusing on deep learning and its basic


constituent parts,functionality ,applications and its coding areas which
will be useful in current time frame.

4. In recent time AI has witnessed explosive growth and in almost every


field there is an application of AI and when we specify it,its main product
is machine learning.

5. Frequently used algorithm in machine learning are Artificial Neural


Networks and they are called deep learning based frameworks.

6. In this course we are going to focus on what are basic deep learning
parts and how can we understand them.Study their Tool kits and how
can we apply these tool kits in our applications.

7. Lets start from the basic definition,Machine learning is defined as it is


the field of study that gives computer the ability to learn without being
explicitly program.

8. This enable a computer to perform a certain task without coding.


9. Computer start learning from its experiences instead of learning.

10. Computer completes its task without human intervention and


explicitly coding.

11. Now, to train a computer we need training data or training set to train
our computer and by using this training data computer learn to perform
its next task.

12. If I want to make an application like face recognition,then first I


have to taught my computer face shape and location.

13. In face detection we give computer 100 images and we draw


bounding boxes against each face and train our computer to detect the
face.

14. Now, our computer become trained on 100 images and able to detect
face.

15. From hundred images each image is a training instance and each
training instance have different properties,because each person has its
different face features.

16. For example: A computer program is said to learn from experience E


With respect to some task T and some performance measure P, If its
performance on T, as measured by P,improves with experience P.
 Experience E is training data
 Particular performance measure is called accuracy.
Lecture 1.1

1. What are different options available to play with machine learning


system,What are the methods to implement machine learning algorithm
and what information we needed to make an machine learning algorithm.

2. What is the phenomena,what is the mechanism and what is the


methodology behind these machine learning applications.

3. The methodology is divided into 3 types


 Supervised or unsupervised learning
 Online versus batch learning
 Instance-based versus model_based learning

4. Common phrase about machine learning is that what is the type of


learning mechanism.

5. The important step in machine learning is that system should learn


automatically from data.

6. The process can be of two types either supervised or unsupervised.

7. In supervised learning there is a labeled data .

8. In Unsupervised learning we have data but labels are not available.

9. In offline learning,first we collect data and then we train our computer.


10. In online learning,we trained our system on available data and with
new data system is automatically updated.

11. We have data of different parameters.Data always available in


spreadsheet,excel sheet and in notepad.

12. In instance base learning is instance base , we tell value against each
row.
13. Model based is more important than instance base learning , we take
whole data and trained the whole model and that model predict the future
values.

Lecture 1.3
1. The best method to differentiate is that how they are learning with their
capabilities.
2. In supervised learning,we train our computer with labeled date same
as we taught a lesson to a children.
3. There are four mechanisms for training our model
 Supervised learning
 Unsupervised learning
 Semi_supervised learning
 Reinforcement learning

4. In supervised learning, the training data is set into the algorithm,which


also includes the algorithm.
Numbers Labels
1- One
2- Two
3- Three
5. Now ,if we change the orientation of 1,then its remain one in different
orientation.
6. In supervised learning ,we train our computer with all the orientations
of the data.
7. The famous task in supervised learning is called classification.
8. To predict the class of a given data is called classification.
9. The classification may be binary or multi-labeled.
10. In binary classification,suppose we have an email system an email
either be a spam email or unspam email.

11. A human is gender wise mail or female,so this is binary label or


binary classification.
12. In multi-label,there are more than 2 labels .
13. Classification is the most important application of supervised
learning.
14. Regression is also a type of classification,but in regression the label
would be in the form of numbers or any numeric value I.e the age of a
children or price of any property.

Lecture 1.3

1. In Supervised learning we train our system with labeled data and then
after training the system is able to predict the future values.

2. unsupervised means that there is no data with labels.

3. Now, the question is how to learn by raw data.


4. The most common method for learning in unsupervised learning is
clustering,clustering means making group of data on the basis of
colour,shape and size.

5. Like all orientation of B in one cluster and the other cluster have all
orientation of alphabet A.

6. System automatically retrieve the data and find instances which are
closer to each other and group them in one cluster.

7. Another algorithm for learning or training process in supervised


learning is Visualization and dimensionality reduction,visualization
means your instances,variables and parameters are visualized.

8. In visualization we have many methods like dots plot in which we plot


x and y coordinates of data and check the relationship between them
I.e,calculate the distance between two variables.

9. In dimensionality reduction we extract the main and important points


from whole data which are used to solve machine learning algorithm.

10. For example in human face recognition application,for face


recognition we have different features I.e
forehead,eyes,nose,lips,complete mouth ,teeth,chin,ears etc,now question
is what is the most important feature.

11. Dimensionality reduction automatically detect the important features


according to our current data.In this technique by using less features we
solve our problem.

12. Association Rule learning describes the features


relationship ,common example of this is market basket analysis.
13. For Example ,if you buy some eggs,then there is more chances that
you will purchase eggs with it rather than purchasing onions and potatoes.

Lecture 1.4

1. The famous approach of machine learning is reinforcement learning.

2. In reinforcement learning ,if our system performance is good,then we


give reward and if system or model does not perform properly then give it
penalty.

3. Reward and penalty means that when you apply given data to
algorithm ,classifier or trainer and if its accuracy is increasing or
improving then score it positive.

4. Score means that when we want to predict label,we pick some


parameters which are good and score them to a high value for a positive
result.

5. Due to some parameters our system performance may reduce or


decrease,then we give penalty to these parameters.

6. Now, we know those parameters or features which are helpful for us in


achieving our final label and also we identify those parameter which may
reduce our system performance.

7. In machine learning problem first we observe data then we select


action using policy and perform our action.
8. After performing action set rewards or penalty, if we choose correct
parameters for prediction then give reward otherwise give penalty.

9. Update the policy according to the results for better performance and
iterate until an optimal policy is found.

10. For example: An agent learn whether it go toward water or fire.If an


agent go towards fire then give penalty (-50 points) and now our system
update its policy that its harmful for an agent to go towards fire.On the
other hand if agent go towards water then we give reward to it.

11. Reinforcement learning works on reward or penalties,if the action is


correct action then it receive reward and if the action is wrong it will
receive penalty.

Lecture 1.6

1. In machine learning algorithms two things are important one is


variables or features that represent our data and the other one is label.

2. One machine learning approach is online learning versus batch


learning or offline versus batch learning.

3. Moat of the machine learning algorithm are based on offline learning,


in offline learning first we collect our data and then apply algorithm on it
and achieve our results.

4. For offline learning, if new data come then we recreate our whole
system for it and then we generate output.
5. In online learning ,you are intact with the system machine when it is
working or processing,in machine learning process means the system
train or learn on the basis of data.Your system will understand the
upcoming data and generate results.

6. Online learning is better than offline learning.

7. But in online process our computation is very expensive because our


system will be active 24 hours for this process i.e, ready to achieve data.

8. In offline learning the system is incapable for learning incrementally


means that it must be trained using all the available data and it takes a lot
of time and computing resources.

9. Offline learning is also called batch learning,but if we want a batch


learning system to know about new data then we need to train anew
version of the system from scratch on the full data set.

10. In online learning, we train our algorithm using data,evaluate


solution ,launch and analyse error.

Potrebbero piacerti anche