Machine Learning: Introduction Sattiraju Prabhakar CS898O: Lecture#1 Wichita State University 1
Topics What is Machine Learning? Why Machine Learning is Important? What do you need to learnt? How do we go about learning it in this course: Course Organization Classes and Labs Evaluation 8/19/2004 ML2004_Introduction 2 2
What is Machine Learning? 8/19/2004 ML2004_Introduction 3 3
Machines Without Learning Machine Interactions E N V I R O N M E N T Task Performance 8/19/2004 ML2004_Introduction 4 Task: It is a class of problems. Example: Diagnosis, Navigate, Design, Explain, Play a game Problem: A specific world situation which needs to be transformed Example: Diagnosing a disease given a set of specific symptoms Example: Navigating from your chair to the blackboard Example: Designing a computer program that does insertion sort Environment: Environments have inputs and outputs. They are models of external world in which the machine learns. Environments are the source of experiences. Interactions: Different kinds of interactions are possible. When there is no learning, the machine addresses a task in environment. We assume the machine already knows how to address a problem Task Performance: The machine is able to address the task in environment and gets desirable results. Results are said to be desirable if they meet some criterion. Please note that the interactions of machines make no sense if they do not address a task and some performance criterion. 4
Example: Diagnostic System Environment: Patients and Hospital Task: Inputs: Symptoms and Lab Tests Outputs: Identification of diseases Therapy recommendations Performance: Ratio of successful therapeutic response to failure response 8/19/2004 ML2004_Introduction 5 This slide indicates where there is no machine learning. But the machine is able to show some performance with respect to task. Problem: Specific symptoms and lab tests Solution: Identity of disease and therapy recommendation Task: The class of all such problems Performance: Shows clearly the successes over the failures. Good performance indicates larger number of successes compared to failures. Poor performance indicates dominance of failures. 5
Example: Programming System Environment: Programming environment that can compile and execute code Task: Input: Code Output: Errors and results Performance: The ratio of successful output results to errors 8/19/2004 ML2004_Introduction 6 Similar to previous slide 6
Example: Engineering Design Environment: Device simulation system Task: Inputs: Functional Specification of Device Specification of device working environment Outputs: Structure and behavior of device Performance: The ratio of correct designs to faulty designs 8/19/2004 ML2004_Introduction 7 The task here is to generate devices given a description of what the devices are supposed to do. The performance of task here is the ratio of successful designs to faulty designs. 7
Machine Learning Problem Model Model Model Model Model Learning Algorithm Machine Task Interactions Learning Interactions E N V I R O N M E N T Emerging Performance 8/19/2004 ML2004_Introduction 8 In this slide, a machine that learns is shown. There are two kinds of interactions with the environment. Learning Interactions: The machine receives experiences from the environment. Using the learning algorithm, the machine builds a model (later we call this target function) for the task. Task Interaction: Using the target function or model, the machine is able to perform tasks in the environment. Performance Improvement: The performance f the machine is improved due to the learning of target function. When there was no target function, the machine could not have shown good performance. With learning that performance improves. 8
Formal Definition of Learning A computer program is said to learn from experience E, with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E. 8/19/2004 ML2004_Introduction 9 This is a very important formal definition of learning. The learning interaction in previous slide gives the machine experience. The machine knows that it needs to perform a number of tasks. It also knows the performance measure. It uses these to improve its performance. See next slide to see how such an improvement is possible. 9
Elements of Learning System E N V I R O N M E N T Task Execution Training Examples Target Function Learning Algorithm Learning System 8/19/2004 ML2004_Introduction 10 The machine has some experiences with the environment. These experiences are the training examples given to the learning system. For these training examples, the learning system computes a target function. The Learning system uses the target function t perform tasks in the environment. The target function is very essential. Without it, the tas performance improvement cannot be done. 10
Examples Learning from Patient Records: Machine learns to diagnose after looking at several past patient records Machine learns to recommend therapy after seeing response of several patients Credit Risk Analysis: Machine learns to predict credit risk after looking at several credit records Learning to Navigate: Robot navigates through corridor We provide several episodes of robot navigating towards a goal as successes The robot learns to navigate towards the goal 8/19/2004 ML2004_Introduction 11 For each of these examples, identify the training examples. The target function is not obvious. In the coming classes we will examine a number of target functions. 11
Example: Handwriting Recognition T: Recognizing and classifying handwritten words within images P: Percept of words correctly classified E: a database of handwritten words with given classifications 8/19/2004 ML2004_Introduction 12 12
Example: Robot driving learning problem T: Driving on public four-lane highways using vision sensors P: Average distance traveled before an error (as judged by a human overseer) E: A sequence of images and steering commands recorded while observing a human driver 8/19/2004 ML2004_Introduction 13 13
Why Machine Learning is Important? 8/19/2004 ML2004_Introduction 14 14
Applications: Data Mining Data mining is finding abstract patterns on all data Examples: Business Applications Web Interactions Knowledge Discovery 8/19/2004 ML2004_Introduction 15 15
Applications: Text Recognition Text Learning: http://www-2.cs.cmu.edu/~textlearning/ 8/19/2004 ML2004_Introduction 16 16
Applications: Knowledge Discovery In Knowledge based Systems We do not need to explicitly put in all the knowledge Resource 8/19/2004 ML2004_Introduction 17 17
What do you need to Learn? 8/19/2004 ML2004_Introduction 18 18
Machine Learning =??? Main aspect is algorithms All these algorithms are characterized by: Learning a Target function from a set of training examples to satisfy a performance criterion Applications require collection of large amounts of data Knowledge representation is an important aspect Implementation and analysis of algorithms 8/19/2004 ML2004_Introduction 19 19
About This Course 8/19/2004 ML2004_Introduction 20 20
Course - Aspects Course Organization Class Room Instruction Labs Evaluation 8/19/2004 ML2004_Introduction 21 21
Course Organization Mid-semester point Symbolic Learning Numerical Learning 8/19/2004 ML2004_Introduction 22 22
Class Instruction Main Goals: Understanding Algorithms Understanding implementation of algorithms Analyzing algorithms Steps in Understanding Algorithms: Visual Understanding Seeing some examples Functional modules of algorithms Expanding functional modules to understand the actual algorithm Working with some examples Analysis of algorithms 8/19/2004 ML2004_Introduction 23 23
Labs Java Instruction Implementation of Algorithms Working with Weka 8/19/2004 ML2004_Introduction 24 24
Evaluation Exams:4 Goal: Understanding of algorithms Assignments: Max 4 Solving problems by applying algorithms Programming Assignments: Max 4 To apply given code to new data To modify given code to develop a code for an algorithm Quizzes: Several Surprise: class and lab 8/19/2004 ML2004_Introduction 25 25