Lect 8 learning types (M.L.).pdf
- 1. Computers Eng. & systems students By Prof. Amira Yassien
- 3. 1. Supervised learning a. Regression Example: Learning to predict houses’ prices Suppose you collect a dataset of houses’ prices in a certain geographical area. Suppose you collect statistics about how much houses cost according to the square footage (feet2) of the house. x x $ x x x x feet2 Cost will be If my house is here x x x x
- 4. x x $ x x x x feet2 Cost will be If my house is here x x x x
- 5. The reason for calling this a supervised problem is that we provide the algorithm a dataset of a punch of houses’ sizes and actual prizes. We simply supervise the algorithm, we give the algorithm the quit right answer for the prices and we want the algorithm to learn the association between the i/ps and o/ps so it gives us more about the right answers. This was an example of what is called a REGRESSION problem. The term regression reverse the fact that the o/p you are trying to predict is a continuous value of the price.
- 6. 1. Supervised learning b.classification Another case of supervised learning problem is the classification in which the variable you are trying to predict is discrete rather than continuous. Example: Let's say you collect a dataset on a cancer tumor and you want the program o/p to predict whether the certain tumor is malignant (harmful bad tumor) or benign. So we collect a number of features of that tumor for the sack of having a simple explanation.
- 7. +1 _ Malignant x x x x x x x x x Tumor size Variables here take values either 0 or 1. The features we collect here is the size of the tumor, depending on the size of the tumor we are trying to decide whether the tumor malignant or not. The Y axis is either 0 or 1 so the dataset may look like this:
- 8. Example: More generally there's other examples that use more than 1 i/p variable (multi-i/p) to try to predict whether the tumor is malignant or not. So patients of different ages having tumors of different sizes will represents the new dataset. age x x x x o x x x x x o o o o o o o o oo o Tumor size o malignant x benign
- 9. You want a program to help you predict whether the new patient of specific age and having certain tumor size whether he has malignant or benign tumor. One program may suggest to separate the 2 classes of the tumors using a straight line(figure) So if we have a patient that lie as figure (green point) so u may classify the patient tumor as benign rather than malignant. So we could predict the tumor type based on 2 i/ps. There are other features (other medical properties ) that the program can use regarding the properties of the tumor like thickness, uniformity of cell size, and uniformity of cell shapes, adhesion and so on.
- 10. One of the most interesting things that M.L. is concerned with is what if your data doesn't lie in a two-dimensional or three-dimensional or sort of even a finite dimensional space. What if your data actually lies in an infinite dimensional space? Our plots here are two-dimensional space.
- 11. Support Vector Machines SVM SVM actually takes data and maps data to an infinite dimensional space and then does classification using not only two features like we've done, but an infinite number of features. And so it turns out that one of the most successful classes of machine learning algorithms: Support Vector Machines
- 12. Learning theory we want to explain how learning programs work so we can apply these learning programs as effective as possible. We need to know How much training data we need in order for your learning program to recognize the patterns between the square foot inch of the house (size of the house) and its price. If you are trying to design a learning program you should spend more time collecting the data. In Machine Learning we need to know how to take the tool of machine learning and really apply them and learn how to use them well.
- 13. Supervised learning An example of a vehicle known as ALVINN. It was done using algorithm called Neural Network. The essential learning algorithm for this is called gradient descent, This is supervised learning because it's learning from a human driver, in which a human driver shows that we're on this segment of the road, I will steer at this angle. This segment of the road, I'll steer at this angle. And so the human provides the number of "correct" steering directions to the car, and then it's the job of the car to try to learn to produce more of these "correct" steering directions that keeps the car on the road.
- 14. And as the algorithm collects more examples and learns of a time, it starts to more confidently choose a steering direction. ALVINN was an example on supervised learning, and in particular it was a regression problem, because the vehicle is trying to predict a continuous value variables of a continuous value steering directions.
- 15. Practical Application The ALVINN ( Autonomous Land Vehile In a Neural Network) project. The ALVINN takes road images from a camera and a laser range finder as an input (video and range information). It produces as o/p the direction the vehicle should travel in order to follow the road. The NN performed much better after it had been trained to recover after making mistakes during driving.
- 17. Unsupervised learning In supervised learning the data you have gives you the right answer to your patient and you want the program just to produce more the same. In contrast, in unsupervised learning problem, you are given a dataset and I’m not going to tell you what the right answer is, just want you to find some interesting structure in this data.
- 18. Clustering would be one example of unsupervised learning, as the program finds structure in the data in the form of data are being partitioned into e.g. two clusters. Unsupervised learning is also used in many problems, For example applying an unsupervised learning algorithm for trying to understand genes data, trying to look in genes in individuals and group them into clusters based on properties of how do genes respond to different experiments. t x x x x x x x x x x xx
- 19. Another Application of unsupervised learning is Image processing. If you give this sort of data, say an image, to a certain unsupervised learning algorithm they will then learn to group pixels together and say those sort of pixel seem to belong together and the images you see in the bottom are grouping to what the program has done to group different pixels together, This sort of clustering algorithm (unsupervised learning algorithm) which learns to group pixels together tends out to be useful for many application in computer vision image processing.
- 20. An example of unsupervised learning Imagine a party and lots of people standing all over , everyone is talking, You find it so difficult hearing a person talking in front of you. the problem here is that with everybody talking can you separate voices of persons you are talking with? Let that the party is attended by two people and we put 2 microphones in a room, microphones are at a slightly different distances and 2 people speak in a slightly different volumes,
- 21. each microphone will pick an overlapping combination of these 2 people's voices so slightly different overlapping combination of voices so speaker 1 maybe over loud on micro 1 and speaker 2 maybe over loud on micro 2, The question is: given these microphone recording can we separate the 2 voices ? MIC 1 MIC 2 Speaker 1 Speaker 2
- 22. Conclusions Unsupervised learning doesn't give the right answer however we ask the program what structure is there in these data. You will discover that this data can actually be explained by 2 independent speakers speaking at the same time and can further separate those 2 voices. That learning algorithm sounds to be so complicated, to take these overlapping audio stream and separate them out? But how complicated is it???? To implement this ( one line of code in MATLAB) several of lines and pages by C or JAVA
- 23. This specific unsupervised learning program are also applied to a punch of other text processing or understanding functional grading and machine data problems such as, the magneto- encephalogram would be an EEG data.
- 24. Reinforcement learning This refers to problems where you don't do one shot decision-making. So, for example, in the supervised learning cancer prediction problem, you have a patient come in, you predict that the cancer is malignant or benign. Then based on your prediction, maybe the patient lives or dies So you make a decision and then there's a consequence. You either got it right or wrong. In reinforcement learning problems, you are usually asked to make a sequence of decisions over time.
- 25. So reinforcement learning is — I think of it as a way for you to specify what you want to be done, so you have to specify when to “reward" and when to “punish". And then it's up to the learning algorithm to figure out how to maximize the reward signals and minimize the punishments. Reinforcement learning is applied to other problems in robotics.
- 26. It's applied in web crawling. A punch of other problems in Robotics using reinforcement learning : 1. A four legged robotic dog 2. A robotic snakes: using learning algorithm to teach the snake to climb obstacle 3. Use learning algorithm to teach a car how to drive in high speeds off road avoiding obstacles 4. Teach a robot how to get over an obstacle All of these robots are difficult to hand code controller for but Using these source of learning algorithm in its short order you can get a robot to do amazing things
- 27. Web crawling and reinforcement learning Amount of accessible information on World Wide Web is increasing rapidly, so that a general-purpose search engine cannot index everything on the Web. Focused crawlers have been proposed as a potential approach to overcome the coverage problem of search engines by limiting the domain of concentration of them. Focused crawling is a technique which is able to crawl particular topical portions of the WWW quickly and efficiently by following only the most interesting links and not having to explore all Web pages. e.g., "crawl pages about baseball", or "crawl pages with large PageRank"
- 29. A Web crawler, sometimes called a spider, is an Internet bot that systematically browses the www, typically for the purpose of Web indexing (web spidering). Web search engines and some other sites use Web crawling or spidering software to update their web content or indices of others sites' web content. Web crawlers can copy all the pages they visit for later processing by a search engine which indexes the downloaded pages so the users can search much more efficiently.
- 30. Learning paradigm supervised: 1- Learning with a Teacher. 2- fast but unnatural regression classification unsupervised 1-no external teacher or critic to oversee the learning process. 2- slow, natural clustering Reinforcement continued interaction with the environment in order to minimize a scalar index of performance 1- Credit assignment problem 2-Delay response.