CE213 Artificial Intelligence Lecture 13

Similar documents
Perceptron. (c) Marcin Sydow. Summary. Perceptron

Simple Neural Nets for Pattern Classification: McCulloch-Pitts Threshold Logic CS 5870

CE213 Artificial Intelligence Lecture 14

ARTIFICIAL INTELLIGENCE. Artificial Neural Networks

Artificial Neural Networks" and Nonparametric Methods" CMPSCI 383 Nov 17, 2011!

Lecture 7 Artificial neural networks: Supervised learning

Multi-Dimensional Neural Networks: Unified Theory

Reification of Boolean Logic

SPSS, University of Texas at Arlington. Topics in Machine Learning-EE 5359 Neural Networks

COMP9444 Neural Networks and Deep Learning 2. Perceptrons. COMP9444 c Alan Blair, 2017

Lab 5: 16 th April Exercises on Neural Networks

Neural Networks. Fundamentals of Neural Networks : Architectures, Algorithms and Applications. L, Fausett, 1994

Neural Networks Introduction CIS 32

ARTIFICIAL NEURAL NETWORK PART I HANIEH BORHANAZAD

Artificial Neural Networks The Introduction

Networks of McCulloch-Pitts Neurons

Neural Networks (Part 1) Goals for the lecture

2015 Todd Neller. A.I.M.A. text figures 1995 Prentice Hall. Used by permission. Neural Networks. Todd W. Neller

Neural Networks. Chapter 18, Section 7. TB Artificial Intelligence. Slides from AIMA 1/ 21

Single layer NN. Neuron Model

Computational Intelligence

Week 3: More Processing, Bits and Bytes. Blinky, Logic Gates, Digital Representations, Huffman Coding. Natural Language and Dialogue Systems Lab

Neural Networks. CSE 6363 Machine Learning Vassilis Athitsos Computer Science and Engineering Department University of Texas at Arlington

Linear Classifiers: Expressiveness

Computational Intelligence Winter Term 2009/10

Machine Learning. Neural Networks. (slides from Domingos, Pardo, others)

Artificial Neural Networks. Part 2

Chapter 9: The Perceptron

Introduction to Neural Networks

Computational Intelligence

COGS Q250 Fall Homework 7: Learning in Neural Networks Due: 9:00am, Friday 2nd November.

EEE 241: Linear Systems

Artificial Neural Networks. Historical description

Linear Regression, Neural Networks, etc.

Machine Learning. Neural Networks. (slides from Domingos, Pardo, others)

In the Name of God. Lecture 11: Single Layer Perceptrons

XOR - XNOR Gates. The graphic symbol and truth table of XOR gate is shown in the figure.

Neural networks. Chapter 19, Sections 1 5 1

Machine Learning. Neural Networks. (slides from Domingos, Pardo, others)

Lecture 4: Perceptrons and Multilayer Perceptrons

Introduction To Artificial Neural Networks

Principles of Artificial Intelligence Fall 2005 Handout #7 Perceptrons

Part 8: Neural Networks

AN INTRODUCTION TO NEURAL NETWORKS. Scott Kuindersma November 12, 2009

Simple Neural Nets For Pattern Classification

Neural Networks Introduction

COMP304 Introduction to Neural Networks based on slides by:

) (d o f. For the previous layer in a neural network (just the rightmost layer if a single neuron), the required update equation is: 2.

Sections 18.6 and 18.7 Artificial Neural Networks

Neural networks. Chapter 20. Chapter 20 1

Multilayer Perceptron Tutorial

The perceptron learning algorithm is one of the first procedures proposed for learning in neural network models and is mostly credited to Rosenblatt.

CSC Neural Networks. Perceptron Learning Rule

Artificial Neural Networks Examination, June 2004

Artificial Neural Network and Fuzzy Logic

Lecture 4: Feed Forward Neural Networks

Introduction to Artificial Neural Networks

Artificial neural networks

Neural Networks: Basics. Darrell Whitley Colorado State University

Artificial Neural Networks Examination, June 2005

Single Layer Perceptron Networks

22c145-Fall 01: Neural Networks. Neural Networks. Readings: Chapter 19 of Russell & Norvig. Cesare Tinelli 1

Artificial Neural Networks Examination, March 2004

CMSC 421: Neural Computation. Applications of Neural Networks

DM534 - Introduction to Computer Science

Neural Networks for Machine Learning. Lecture 2a An overview of the main types of neural network architecture

The Perceptron. Volker Tresp Summer 2014

CSE 352 (AI) LECTURE NOTES Professor Anita Wasilewska. NEURAL NETWORKS Learning

Unit 8: Introduction to neural networks. Perceptrons

CS:4420 Artificial Intelligence

Artificial Neural Networks. Q550: Models in Cognitive Science Lecture 5

y k = (a)synaptic f(x j ) link linear i/p o/p relation (b) Activation link linear i/p o/p relation

Last update: October 26, Neural networks. CMSC 421: Section Dana Nau

Ch.8 Neural Networks

Sections 18.6 and 18.7 Artificial Neural Networks

EE04 804(B) Soft Computing Ver. 1.2 Class 2. Neural Networks - I Feb 23, Sasidharan Sreedharan

Input layer. Weight matrix [ ] Output layer

Vote. Vote on timing for night section: Option 1 (what we have now) Option 2. Lecture, 6:10-7:50 25 minute dinner break Tutorial, 8:15-9

Artificial Intelligence (AI) Common AI Methods. Training. Signals to Perceptrons. Artificial Neural Networks (ANN) Artificial Intelligence

Neural networks. Chapter 20, Section 5 1

ECE521 Lectures 9 Fully Connected Neural Networks

Artificial Neural Networks

Revision: Neural Network

Symbolic vs. subsymbolic representation in cognitive science and artificial intelligence Vladimír Kvasnička FIIT STU.

Pattern Recognition Prof. P. S. Sastry Department of Electronics and Communication Engineering Indian Institute of Science, Bangalore

Multilayer Perceptron

Finite Automata. Warren McCulloch ( ) and Walter Pitts ( )

BOOLEAN ALGEBRA INTRODUCTION SUBSETS

AI Programming CS F-20 Neural Networks

XI STANDARD [ COMPUTER SCIENCE ] 5 MARKS STUDY MATERIAL.

Intelligent Systems Discriminative Learning, Neural Networks

CHALMERS, GÖTEBORGS UNIVERSITET. EXAM for ARTIFICIAL NEURAL NETWORKS. COURSE CODES: FFR 135, FIM 720 GU, PhD

Answers Machine Learning Exercises 4

Neural Networks. Nicholas Ruozzi University of Texas at Dallas

Practicals 5 : Perceptron

Linear discriminant functions

Sections 18.6 and 18.7 Analysis of Artificial Neural Networks

INTRODUCTION TO NEURAL NETWORKS

Lecture 6. Notes on Linear Algebra. Perceptron

Neural Networks. Xiaojin Zhu Computer Sciences Department University of Wisconsin, Madison. slide 1

Transcription:

CE213 Artificial Intelligence Lecture 13 Neural Networks What is a Neural Network? Why Neural Networks? (New Models and Algorithms for Problem Solving) McCulloch-Pitts Neural Nets Learning Using The Delta Rule Back Propagation Networks 1

McCulloch-Pitts Neural Nets The Basic Unit (artificial neuron, mapping input to output) It is called an MP neuron or MP unit (figure on the left) for simulating a typical biological neuron (figure on the right). x 1 w 1 x 2 x 3 w 2 w 3 θ y w n x n where x i is the ithinput w i is the weight of the ithinput θ is the unit s threshold y is the unit s output A typical biological neuron (Approximately 100 billion neurons in the human brain) 2

McCulloch-Pitts Neural Nets (2) The output (binary) is computed from the input as follows: θ =1 <θ =0 Basic information processing: weighted sum and thresholding, which could be quite powerful (e.g., fault diagnosis). Note that this is much more like a logic gate than a biological neuron. 3

What Can an MP Unit Compute? Assume inputs are binary (i.e., 1 or 0) AND A 1.5 B Y = 1 if A + B 1.5; Y = 0 otherwise. Y Of course, there are an infinite number of other correct designs, e.g., changing to 1.1 or 1.2 would do. A B Y 0 0 0 0 1 0 1 0 0 1 1 1 (Aand Bcorrespond to x 1 and x 2 in the MP neuron model presented in the previous slides. The weighted sum is w A A+w B B, wherew A = w 1 = and w B = w 2 =.) 4

What Can an MP Unit Compute? (2) OR A B 0.5 Y = 1 if A + B 0.5; Y = 0 otherwise. Y A B Y 0 0 0 0 1 1 1 0 1 1 1 1 5

What Can an MP Unit Compute? (3) NOT A - -0.5 Y Y = 0 if A > 0.5; Y = 1 otherwise. 6

What Can an MP Unit Compute? (4) Are there any Boolean functions that an MP unit cannot compute? How about XOR (exclusive OR)? A B A xorb 1 1 0 1 0 1 0 1 1 0 0 0 7

What Can an MP Unit Compute? (5) Suppose weights are w A and w B, input to the MP unit isw A A+w B B. Then we require for performing XOR (check the equations on slide 3): w A + w B < θ w A θ w B θ (1) (for A=1, B=1, Y=0) (2) (for A=1, B=0, Y=1) (3) (for A=0, B=1, Y=1) 0 < θ (4) (for A=0, B=0, Y=0) But clearly if both (2) and (3) are true, then (1) must be false, since (4) implies that θ, w A and w B are positive. Thus there is no combination of weights and thresholds that will create an MP unit that computes the XOR function. An MP unit cannot compute XOR! 8

What Can an MP Unit Compute? (6) A Graphical Interpretation Consider the ORunit: w A = w B = 1; θ= 0.5 What is the range of values for which Y = 1? All points for which w A A+ w B B> θ A 0.5 A + B = 0.5 Y = 1 i.e., whenever A + B > 0.5 Y = 0 0.0 0.0 0.5 Thus: All points above the line A + B = 0.5 lead to Y =1 All points below the line lead to Y = 0 B 9

What Can an MP Unit Compute? (7) Similarly for the ANDunit: w A = w B = 1; θ= 1.5 1.5 A + B = 1.5 A Y = 1 0.5 Y = 0 0.0 All points above line A + B = 1.5 lead to Y =1 All points below the line lead to Y = 0. 0.0 0.5 1.5 B 10

What Can an MP Unit Compute? (8) The General Case for 2 Inputs θ/w A w A A+ w B B= θ x 2 initial weights A Y = 1 Y = 0 0.0 0.0 Y = 1 if and only if w A A+ w B B θ. B θ/w B θ, w A, w B can be determined by learning from examples. w A A+ w B B= θis a straight line that interceptsthe axes at θ/w A and θ/w B. x 1 11

What Can an MP Unit Compute? (9) So what about XOR? A 0.5 0.0 0.0 0.5 B There is no straight line that separates the blue spots (Y=0) from the red spots (Y=1). 12

What Can an MP Unit Compute? (10) More Than 2 Inputs (one unit only) This result generalises to any number of inputs: θ =0 This is the equation of a hyperplane in n-dimensional space. So an MP unit divides the space into two regions separated by a hyperplane. The values of the weights and threshold determine the position and orientation of this hyperplane. This may solve complicated classification problems in high-dimensional space! ( First golden age of neural network research in 1960s) 13

Linear Separability Thus the only functions that an n-input MP unit can compute are those defined by a linear surface: a straight line when n= 2 a plane when n= 3 a hyperplanewhen n> 3 Such functions are said to be linearly separable. For this reason MP neurons are sometimes called: Linear Separation Units(LSUs) (They are not linear units, but linear separation units) 14

More Than One Unit We could obviously use the output of one MP unit as an input to another MP unit. In such a way we could build an artificial neural network. Could such a neural network (more than one neuron) compute functions that a single unit could not? -Yes 15

More Than One Unit (2) XOR with 4 MP units A A B 1.5 B - -0.5 (A B) Is it possible to implement this part by one MP unit only? 1.5 (A B) (A B) A xorb A B A B 0.5 16

More Than One Unit (3) An optional exercise: Devise a 3 unit MP network that computes XOR. Hint: Consider devising an MP unit to compute NOT AND (A B) Designing an MP neural network like what have been done so far seems difficult, especially for high-dimensional input problems. Solution Learning from data/experience (topic for next lecture) 17

Summary McCulloch-Pitts Neural Nets MP Unit (Neuron) What Can an MP Unit Compute? Linear Separability More Than One Unit (Neuron) 18