Artificial intelligence, AI

What ?

definition

Lynne Parker

Artificial intelligence refers to "a broad set of methods, algorithms and technologies that make software 'smart' in a way that may seem human-like to an outside observer," said Lynne Parker

categories

Acting Humanly (ponašati se ljudski)

the turing test approach part of RA

What?

total turing test

capabilities

compoter vision

robotics

history

Alan Turing ( 1950 )

was designed to provide a satisfactory
operational definition of intelligence.

capabilities

natural language processing

knowladge representation

automated reasoning

machine learning

Thinking humanly (razmišljati ljudski)

the cognitive modeling approach

brings together

psychology =>

theory of human mind (cognitive science)

computer model from AI

Thinking rationally (razmišljati racionalno)

the "laws of thought" approach part of RA

brings together => intelligent system

logic

AI

history

Aristotel => "right thinking"

Acting rationally (ponašati se racionalno)

the rational agent(RA) approach

limitet rationality

foundation

Philosophy

Mathematics

Economics

theories

utility

Desision theory

combines probability theory with utility theory

Game theory

Neuroscience

neuron

Psychology

cognitive science

computer engineering

Control theory and cybernetics

control theory

history

Norbert Wiener

Ashby’s Design for a Brain (1948, 1952)

elaborated on his idea that intelligence could be created by the use of homeostatic devices containing appropriate feedback loops to achieve stable adaptive behavior

Linguistics

grew up hybrid field

computational linguistics ( natural language proseccing )

theory

click to edit

methods

Machine intelligence

machine learning

Cognitive computing

it is often a complete architecture of multiple A.I. subsystems that work together

definition

"This is a subset of A.I. that deals with cognitive behaviors we associate with 'thinking' as opposed to perception and motor control," Dietterich

more less => marketing trick

Projects

games

examples

deepmind

theory

super mario bro

PySc2 for StarCraft II

OpenIA

new technic of learning

Evolution Strategies, ES

scalable alternative to reinforcement learning(RL)

optimizatio technique

ML

Titanic Data science

ML from disaster => kaggle data project

machine learning

click to edit

steps

preparing the data

training a model

evaluating the model

improving the performance

What?

covering multiple technogies

methods

artificial neural networks, ANN
#

types

deep Q-network, DQN

What?

DeepMind Tehnologies

deep Q learning or deep reinforcement learning

combination of deep learning and reinforcement learning

ANIFIS

What?

combination of NN and fuzzy logic

Deep NN

What?

includes a large system of neurons arranged in several hidden layers

learning

deep learning

About

learning model

lerning methods

supervised

unsupervised

Data comes without labels. We need to find regularities in the data

density estimation ❓

(according to ACM)

banches of AI

click to edit

computer vision (rač. vid )

patern repognition (prepoznavanje uzorka)

click to edit

Understanding and processing of natural and artificial languages

natural language procesing ,NLP

components

NL understanding, NLU

NL generation, NLG

involves

text planing

sentence planning

text realization

click to edit

click to edit

tool

Reverse tuning test

fuzzy logic

about

nastaje na pojmu Fuzzy skupova (Fuzzy sets)

klasični skupovi vs fuzzy skupova

fuzzy skupovi

klasični skupovi

jasno određene granice

nisu jasno određene granice

uploaded image

Membership function ( funkcije pripadnosti )

svojstva

Boundary (Granice )

Height ( Visina )

Support ( Osnovica)

Crossover point ( Polovište )

Core (Jezgra)

def. kao klasičan skup svih elemenata za koje je vrijednost veća od 0

def. kao klasičan skup svih elemenata za koje je vrijednost manja od 1

def. kao klasičan skup svih elemenata za koje je vrijednost manja od 1

def. kao klasičan skup svih elemenata za koje je vrijednost maksimalna

def. kao klasičan skup svih elemenata za koje je vrijednost 0.5

type of membership func.

trokut

trapez

Gaussianova krivulja

Generalizirano zvono

koristi

koristi

fazifikacija

agregacija

Fuzzy interence system, FIS (Fuzzy sustav za zaključivanje)

model

input

je sustav za računanje koji raspolaže s više koncepata znanja, a to su :

teorija fuzzy skupova

fuzzy if-then pravila

fuzzy zaključivanje

lingvistička varijabla

output

fuzzy baza podataka

povezana sa :

dio za zaključivanje # ( Agregacija)

defazifikacija

fazifikacija

više modela :

Mamdani

Sugeno

about

pomoću dva parametra donosi zaključak

" riječ " subjektivnog karaktera

if -then pravila => pravila po kojima fuzzy logika donosi zaključke

jezičnog karaktera

sastoje :

premisa

zaključaka

prvi korak, tansformiranje crisp ulaznih parametara u fuzzy ulazne parametre

crisp ulazni parametri

numeričke vrijednosti koje se pretvaraju u odgovarajuće jezične vrijednosti , za to potreban FIU(engl. fuzzy interferance unit)

suprotan fazifikaciji,

sastoje od elemenata koji po nekom kriteriju pripadaju u taj skup

Agregacija

proces kombiniranja fuzzy rezultata pomoću pravila u kojem dobijemo konačan fuzzy rezultat koji se šalje u proces defazifikacije

svodi na logičke operacije konjukcije ili disjunkcije, odnosno T ili S norma


dobiva se koeficijent

metode

centar površine ili centar najviše sume

Center of gravity, COG (centar suma)

prvi, ili srednji ili visinski maximum

visinska defazifikacija

najčešća metoda

What ?

zadaci koje rješava

problem klasifikacija

predviđanje

vezane uz ulazne i izlazne varijable i probleme optimizacije

raspoznavanje uzoraka, obrada slika, govora, simulacija

obrada podataka nepreciznih i nekompletnih podataka

glavna sastavnica

Node, Process unit (Umjetni neuron, procesna jedinica ili čvor)

sastoji

Weight (težina)

težinske sume ili net

funkcije praga ili aktivacijske funkcije

analogna je somi ili tijelu živčane stanice

analogne dendritima

Treshold ( prag ) ili pomak (offset ili bias )

tipovi :

step func.

funkcija praga

sigmoidalna

analogna je integracijskom dijelu neurona

Umjetna neuronska mreža (engl. artificial neural networks, ANN) čini cjelinu međusobno povezanih čvorova čije veze možemo prilagođavati prema skupu podataka za učenje.Paradigma temeljena na biološkoj neuronskoj mreži ili mozgu znatno je pojednostavljena jer se nemogu programirati fenomeni koji posjeduje živčani sustav. Moć obrade podataka nalazi se u vezama između pojedinih procesnih jedinica.

varijable

ulazna ( Signali ) ili funkcijski signal

izlazna

analogija

sinapse kod neurona

about

su realni brojevi u intervalu [-1,1] ili [0,1]

označavanje

x1, x2, xn

analogija

s aksom

označava s y

daje boolov iskaz

f(x) = {

0, net < 0

1, net ≥ 0

f(net) = {

najčešći prijenos funkcije, daje neku prednost kod učenja NN

net , x < net < 0

1, net ≥ x

0, net < x

f(net ) =

net = w0x0 + w1x1 + w2x2 + ... + wnxn

uzimamo θ = - w0

karakteristike koje NN čine različite

broj slojeva procesnih jedinica

prijenosne funkcije

tip veza između ulaznih i izlaznih podataka

tip učenja

procesne jedinice u NN mogu biti spojene

u ANN procesne jedinice ( node) smještene su u slojeve

kriterij određuje tip učenja => da li je izlazna vrijednost poznata ili ne ?

dvoslojne

višeslojne

sastoje se od dva sloja

ulazni sloj

izlazni sloj

sastoje se

skriveni slojevi

izlazni sloj

ulazni sloj

može ih biti više

inter-slojna veza

intra-slojna veza

za povezivanje čvorova u razl. slojevima NN

služe za povezivanje čvorova unutar jednog sloja

primjer Hopfildova mreža

prijenos informacija kroz NN

FeedForward ( Statički )

FeedBack ( Dinamičke )

informacije su plasirane u više slojeve u jednom prolazu i najvažnije informacija se ne vraća u niže slojeve

informacije se vraćaju u niže slojeve,

informacije se mogu plasirati u dva smjera

ulazne informacije primaju u vremenskim intervalima

iz nije poznata

iz poznata

Nenadgledano (Unsupervised )

suprotno nadgledanom učenju => sustav sam otkriva odnose između podataka prema opisanim značajkama

koriste se

prepoznavanje uzoraka

Supervised (nadgledano )

tip učenja kod kojeg sustav sam podešava parametre na temelju podataka s poznatim ulaznim i izlaznim vrijednostima

tipom učenja koristi se za klastriranje ( clustering )

koliko primjera za učenje pokažemo meži za vrijeme trajanja jedne iteracije

Bactch ili off-line (grupno )

Patern-By-Patern or on-line (pojedinačno)

predočavamo sve primjere u jednoj iteraciji, zbog toga NN uči samo u jednoj fazi u kojoj se prilagode težine, a u ostalim fazama težine su fiksirane

NN uči u vremenskim intervalima, u svakom vremenskom intervalu, tj. iteraciji dodjeljen je jedan primjer za učenje NN pritom se vrše prilagodbe težinskih faktora

faze izrade NN :

  1. faza učenja ili treniranja
  1. faza testiranja

neuronsku mrežu nemožemo učiti ako nismo definirali

tip NN

model ( ulazno - izlazne varijable )

karakteristike

podjela prikupljenih podataka u podskupove #

Kako dugo trenirati neuronsku mrežu ?

proces ocjenjivanja mreže

u ovoj fazi težine su fiksirane, te ocjenjujemo kako će se mreža ponašati s novim podacima

diff learning method

Supervised

Adeline

Percepton

Multilayer percepton, MLP (Višeslojni percepton )

sastavljena od percepton čvorova ( nodes )

percepton rule ( pravilo perceptona)

ukoliko se uzorak klasificira ispravno, ne radi korekciju.

ukoliko se uzorak klasificira neispravno, primjeni LMS pravilo

Ciklički uzimaj sve uzorke redom, a postupak zaustavi kada sve uzorke klasificiraš ispravno za redom

temelji na matematičkom modelu McCulloch-Pittsa, poboljšao ga je Frank Rosenblatt

karakteristike

gl. karakteristika

raspoređenih u

više slojeva

izlazni

ulazni

potpuna povezanost njegovih čvorova

signali

vrste

funkcijski signal ili ulazni signal

izlasni signal ( Signal greške )

znači da svaki čvor u svakom sloju povezan sa svim čvorovima predhodnog sloja

signali kroz mrežu kreću s lijeva na desno ( unaprijed )

nastaje u izlasnom čvoru i širi se unatrag sloj po sloj po NN pritom izračunava po svakom čvoru neki tip funkcijske greške

nastaje u ulaznomsloju mreže i širi se dalje do izlaznog sloja te završava kao izlazni signal

skriveni

dva sloja

učenje

backpropaganation algorithm

deep learning

linear algebra

collecting data

About

numerical cumputation

numerical cumputation #

click to edit

fuzzy izlazni parametri pretvaraju se u crisp izlazne parametre

StarCraft AI

agent

what

an intelligent program that acts autonomously in an enviroment

solving several difficult problems

planning

optimization

multiagent control

scene analysis (analiza scena)

click to edit

Accociation of computing Machinery, ACM #

formalisms and methods for knowledge representation

Machine learning

deduction and theorem proving

click to edit

automated programing

Problem solving

Expert systems

Robotics

General AI

cognitive modeling

philosophical foundation

click to edit

Distributed AI (raspodjeljena UI)

Tuning test

history

Alan Tuning

Can machines think?

experiment

The imitation game

Completely Automated Public Turing Test to Tell Computers and Humans Apart, CAPTCHA

completely automated public turing test to tell computes and humans apart

About

click to edit

state space search (pretraživanje prostora stanja)

mathematic model

search problem

consists :

goal state or many goal states

initial state

Screen Shot 2018-02-28 at 12.31.51 PM

transitions between states

Screen Shot 2018-02-28 at 12.33.56 PM

successor function

Screen Shot 2018-02-28 at 12.32.50 PM

test predicate to check if a state is a goal state

can be defined either explicitly

methods

search tree

defining the state transitions

click to edit

control methods ❓

S - set of states (state space)

(basic idea of state space search)

directed graph or digraph (usmjereni graf)

If we also need the transition costs

problems

puzzle

grap vs state space

graph node = states

arcs (directed edges) = transitions between states

may contain cycles

weighted directed graph

By searching through a digraph, we gradually construct a search tree

expanding one node after the other (pojedine čvorove proširujemo)

type of nodes

closed nodes

open nodes or front

nodes that have been generated, but have not yet been expanded

already expanded nodes

using

successor function (operator)

generate the descendants of each node

tree vs state space

Search tree is created by searching through the state space

Search tree can be infinite even if the state space is finite
NB: state space contains cycles ⇒ search tree is infinite

state vs. node

Node n is a data structure

stores

a state

some sdditional data

node data structure

Screen Shot 2018-02-28 at 1.27.38 PM

d - depth of the node in the search tree

s - state

Screen Shot 2018-02-28 at 1.30.19 PM

Screen Shot 2018-02-28 at 1.31.42 PM Screen Shot 2018-02-28 at 1.31.53 PM

comparing problems vs. algorithms

Problem properties:

Algorithm properties:

search startegies

blind or uninformed search (cro. slijepo pretraživanje)

Heuristic (directed or informed) search (cro. usmjereno pretraživanje)

type

Breadth-first search, BFS

type

Greedy best-first search

A* algorithm

Hill-climbing search

click to edit

click to edit

click to edit

Data is labeled and comes in pairs (input, output)=(x, y). We need to find a mapping f(x) = y

if y is a discrete label(nebrojčana vrijednost)

classification

If y is a number

clustering

novelty/outlier detection

dimensionality reduction

reinforcement learning

Learning an optimal strategy based on trials with a delayed credit

AlhaGo

reinforcement learning

regression

how lear model?

click to edit

learning steps

train

test

About

feature extractor (Ekstraktor značajki)

generates the feature vector x for each data instance

during training the model is fed with data

pairs(x, y)

data instance ⇒ x

labels ⇒ y

prediction

create model

during prediction, the model is presented only data instance x as inputgenerates y as output

problem

overfitting

If the model is too complex, it will adapt to much to the data it has been trained on, but give poor predictions on unseen data

solution #

prepare dataset for learning

cross-validation (unakrsna validacija)

split dataset

prunning of decision trees (podrezivanje stabla odluke)

typically

70% ⇒ training set

30% ⇒ test set

because of complexity

split dataset

30% ⇒ validation set

30% ⇒ test set

40% ⇒ training set

We train the model on the train set, use this model to obtain predictions on the test set, and then compute the accuracy

the accuracy

on the test set indicates how well the model generalizes (the model has previously not seen the data from the test set)

algorithms

Na ̈ıve Bayes Classifier

decision trees

click to edit

matrices

vectors

tensors (cro. tenzor)

deep learning

Information theory concepts

:

cross-entropy

decision trees on basic of maximum information gain

Viterbi algorithm

widely used in NLP and Speech

machine tlanslatation => ENCODER - DECODER

technique

clustering

What?

that involves the grouping of data points

Convolutional NN, CNN

used for

effective mechnism used for image recognition

idea

inspired by brain

visual cortex

visual neuron cells in the brain that each act in a different way

simple cells (S cells)

complex cells ( C cells)

What?

activate, for example, when they identify basic shapes as lines in a fixed area and a specific angle

continue to respond to a certain stimulus, even though its absolute position on the retina changes

model

hierarchical NN model

neocognitron

click to edit

history

first

LeNet-5

arhitecture

layers

organised in 3 dimensions

height - H

dept - D

width -W

components

1) the hidden layer/ Feature extraction part

2) the classification part

click to edit

click to edit

scalars

About

What?

About

what?

is an array of numbers, numbers are arranged in order

WHAT ?

2D array of numbers, , so each element is identified by two indices instead of just one

What?

In some cases we will need an array with more than two axes

click to edit

is just a single number

probability & information theory

probability

uncertainty

possible source of uncertainety

2.

3.

1.

Inherent stochasticity in the system being modeled

Incomplete observability

Incomplete modeling

click to edit

click to edit

click to edit

softmax function

Screen Shot 2018-06-27 at 13.01.57

often used to predict the probabilities associated with a multinoulli distribution

numerical computation

poor conditioning

Gradient-Based Optimization

an example of a 0th-order tensor

Screen Shot 2018-06-28 at 08.20.50

built-in scalar types

int

float

complex

bytes

Unicode ( in Python )

an example of 1st-order tensor

an example of 2nd-order tensors

Screen Shot 2018-06-28 at 08.27.35

chatbots

About

an AI software that can simulate a conversation (or chat) with a user in natural language through messaging applications, websites, mobile apps or through the telephone

there are two different tasks at the core of a chatbot


user request analysis

returning the response

2019 ...

Watson Assistant

Bold360

Rulai

LivePerson

Inbenta

Ada

Vergic