Introduction to Artificial Intelligence

Artificial Intelligence (AI)

• AI refers to machines mimicking human cognitive functions.

Birth of AI: 1956

• The Dartmouth workshop in 1956 marked the beginning of AI
• Allen Newell, Herbert A. Simon, and Cliff Shaw introduced the logic theorist.
• 1951: Christopher Strachey created a checkers program and Dietrich Prinz wrote a chess program utilizing the Ferranti Mark 1 machine at the University of Manchester

Overview

• AI systems generally operate through a process of input, algorithm, and output.

Machine Learning

• Machine learning focuses on algorithms that learn and predict from data.
• The machine learning process involves training a model with training data
• Testing the model with test data, and outputs some result

Terminology

• A model represents reality in a simplified way for a specific purpose.
• Predictive models estimate the unknown "target" value using formulas or logical statements.
• Prediction involves estimating an unknown value, also known as the target.
• An instance/example represents a specific fact or data point.
• Models are described by attributes (fields, columns, variables, or features)
• Model induction involves creating models from data.
• Training data serves as input for induction algorithms.

What is a Model?

• Models are simplified representations of reality, based on assumptions
• Examples of models include maps, prototypes, and the Black-Scholes model.
• Data mining examples include formulas for predicting customer attrition

Feature Types

• Numeric features have an order (numbers, dates) and a dimension of 1.
• Categorial features do not have an order (text, names, ratings), and their dimension is the number of possible values

Common Machine Learning Tasks

• Classification and class probability estimation determine the likelihood of a specific outcome
• Regression predicts numerical values.
• Similarity matching identifies similar items or data points
• Clustering groups similar consumers into natural groups.
• Co-occurrence grouping and association rules, also known as frequent itemset mining, identify items commonly purchased together.
• Profiling describes typical behavior to detect fraud.
• Data reduction identifies key dimensions describing consumer preferences
• Link prediction suggests new connections based on existing relationships
• Casual modeling identifies events/actions that influence others

Classification Models

• A decision tree serves as a classification model.

Machine Learning

• Machine learning involves two types of learning: supervised and unsupervised.
• Supervised learning infers a function from labeled training data
• Unsupervised learning infers hidden structure from unlabeled data

Why Deep Learning?

• Deep learning was highlighted by an AI's victory over top Dota 2 players.
• An Elon Musk-funded AI team defeated human players in Dota 2

What is Deep Learning?

• Deep learning employs neural networks with multiple hidden layers
• Artificial neural networks rely on computational models based on simple neural units.

Easy for Me, Hard for You

• Computers excel at calculations and quickly processing basic instructions.
• Image recognition is challenging for computers but becomes achievable with AI.
• Artificial intelligence aims to create algorithms that solve hard problems.

Human vs Computers

Image recognition requires human intelligence that machines currently lack

AI in Simple Words

• AI can be thought of as a machine converting kilometers to miles
• The precise conversion formula is unknown but assumed to be linear.
• The relationship is expressed as miles = kilometers x c

Learning Example

• Kilometers: 100 → miles = kilometres x 0.5 → miles: 50
• Calculates an initial conversion.
• The result is compared to the correct answer
• Error: 12.137.
• The error guides subsequent guesses
• Let's nudge c up from 0.5 to 0.6 and see what happens Now the error is much smaller: 2.137 Kilometres: 100 → miles = kilometres x 0.6 → miles: 60
• The important point here is that the error to guide how the value of c is nudged
• Let's nudge the value of c up again from 0.6 to 0.7
• Kilometres: 100 miles = kilometres x 0.7→ miles: 70 Error: -7.863

Neural Network Learning

• The core process of learning in neural networks involves iteratively improving answer accuracy.
• Machines are trained to get better and better at giving the right answer.
• The term “iterative” indicates repeatedly improving an answer bit by bit.

Quiz

• Refining the model through parameter adjustment based on the difference between the model's output and known true examples.

Classification vs Prediction

• You can clearly see two groups

Training a Simple Classifier

• The goal is to train a linear classifier to correctly classify bugs as ladybirds or caterpillars
• Training examples are needed to learn from
• (Examples of truth that are used to teach a predictor or a classifier are called the training data
• The general fomula fro the line is: y = Ax + B
• A controls the slope
• For simplicity, let's assume that B = 0

Training Example One

• Given y = 0.25 * x as the model
• The with is 3.0 and length is 1.0 for a ladybird. In we tested the modl with this example where width is 3.0, y = 0.25 * 3.0 = 0.75 Too small it needs to be a length of 1.0
• The line needs to be a dividing line between ladybirds and caterpillars, not a predictor of bugs length given its width
• Error = (desired target - achutal output)

Train the Model Based on the Error

• What do we do wih this E to guide us to a better-refined parameter A?
• To inform the requires change in paramter A Let's call the correct desired vlaue, t for target value
• We wanted to know how to adjust A to improve the slope of thr line so it is a better clsssifier
• E = \triangle A x then \triangle = E/ X
• We can use tje error E to refine the slope A of the classifying ling by an amount \triangle A

Let's Do It

• From the fristm exmaple, the error ws 0.35 and the x was 3.0
• Change the curent A = 0.25 by 0.1167 (A + \triangle A wich is 0.25 + 0.1167 = 0.3367
• Calculte the new y woth this new a is. (1.1)

Training Example Two

• An imprtant in machine learning
• Modeate the update
• Instead of jumpong enthusiastically to each new A, we take a fractin of the changde \triangle A e = L (E/x) L- 0.5
• Moderators is offener called a learning rate We only uodate hald as much as would havr done with moderate.

Example

• L = 0.5 AO = 0,25
• \triangle A

Visalizing the Example

(Visualiing the example)

• We are on the wrong side of the training example.
• You can have pudding only if you have

Boolean Logics Functions

• You can have the our pudding olny is youve eaten your vegeteabled and us you're

Simple Liner Classificer

• Is there more ma,aria degress
• It is poaaible fir a simpker linder l

Lecture

• Neurons, Natures Computing Machines
• Pigeon brans were vistkt nore capacle tkangital computrt

Neuron

• All trasmittab elctirca simgnal
• These signals that you send in your
• How many nerons do we ne

Function

• A simple neurok in the brain is 1

More Thing About Neurons

• Mathmatically there is one such acrivation function that coukd aciveve this effort.

Activation Fuction

• A simple activation function is steep and a common activayion function is sigmond

Sigmoid Function

• y = 1/ 1 * e1
• Isn't as scafry as it is frist lyks
• Rectrifed lukear ubht acrivatoon is a 3-Layer Examole with Marxis Miltipky

Neurons

• Real biolofical nevrops take mnagy

More Artifical Model

• Threeo layers each with threet ari

Neural Netwotk with three layers

The weight 3 -9 -9 is suymly what

Question

You should ne conbnect The l;eaming ocrocess wi;l le

Matrix Resposne

Matrix and neural network (input / + / input

x = W - 1 thrat is W js the matri uxb y = M- 7 esvn uf/

Update Weights in New Network

Belive it if nit Emrabraxe pessinmism Imagine a vefy somoevex landscape What's tge link between this real

Updating Wekghts

• 14.4

The finay pesutl to woth

Modertatinkg Step SKze

Wqje increae x im the oppoj te dorectui

Grediemnt descent

I gradient alse meNn s wu o

error function

L function

Funcuton with Two weights

Erior slope

W22