AI

key concepts

auotnomy

adaptivity

"the study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can be in principle be so precisely described that a machine can be made to simulate it."

how to identify?

Turing: intelligent is to behave is intelligent

john searle: chinese room

problem solving

state space

transitions

costs

the set of possible situations

possible moves between one state and other

associated to each transition

related fields

robotics

data science

machine learning

systems that improve their peformance in a given task with more and more experience or data

Neural networks

deep learning

types

narrow AI

handles one task only

AGI artificial general intelligence

still science fiction

strong AI

like a "mind", genuinely conscious

weak AI

behaves like intelligent

game

game tree

minimax algorithm

heuristics

Dealing with uncertainty

Fuzzy logic

Probability

ability to think of uncertainty as a thing that can be quantified at least in principle

uncertainty is not something that is beyond the scope of rational thinking and discussion

Odds

X:Y = For each X occurrences, you will have Y opposite occurencies

X/Y = you will have X occurrences out of Y total occurrences (including X)

It’s easier to intuitively understand odds instead of percentages

Odd x:y = probability x/(x+y)

Bayes rule/formula

Prior and posterior odds, according to new information available

Likelihood ratio

Probability of the observation in case the event of interest/the probability of the observation in case of no event
ex (1/10)/(9/10]

Posterior odds = likelihood ratio prior odds
ex 5,11:2 = 5,1:2

"naive bayes classification"

classifiier is trained by analysing a set of training data

Despite it’s naïveté, works very well in practice

Types

MNIST dataset
Modified National Institue of Standards and Technology

supervised learning

binary classification problems

unsupervised learnings

data visualizations

grouping data in clusters

reinforcement learning

based on feedback about the decisions taken

based on a number of examples and its correlaction to the correct label

Linear regression

Be careful!

Do not become too confident about the accuracy of predictions

Separate data

Training data

Test data

Dont’t make the rules fit the past data perfectly

The correct answers are available to train/compare

Correct answers aren’t provided

Example: models group data and users define labels (loyalty cards programs)

Generative modeling

Given some data, model can generate more of the same (people’s face from pictures, e.g.)

Filter bubbles #

Nearest neighbor

Applicable when result may be a number, rather than a class

Components

Inout

Output

Intercept

Noise

Coefficients/weights

Logistic regression

Used to turn linear regression outputs into predictions about labels

Limits

Hardness of task

ML method

Amount of training data

Quality of data

mathematical, deep equations

Several layers of simple processing units, connected in a network

Inputs are passed through each one of the layers

Neurons

Stand processors, waiting for signals

Dendrites/axons

Connections/wires between neurons

Key features

Neurons process vast amounts of information simultaneously (instead of CPU)

Data storage (memory) and processing aren’t separated

activation function

done after the liear regression calculation

includes

identity function

step function (ON/OFF)

sigmoid function

= output of linear regression

imitates real neurons

numerical levels of activation

click to edit

implications

future prediction

philip tetlock

foxes: many small ideas

better at predictions

hedgehogs: one big idea

pay attention to carefully justified and balanced information sources

be suspicious about people who keep explaining everything using a single argument

terminator will not come

superintelligence will not evolve from developing narrow AI methods to solve real world problems

singularity: exponential intelligence face growing problems that slows down its progress

real threat = Terminator headlines diverts the attention from the actual problems

societal implications

  1. algorithmic bias

caused by human bias in data

human discrimination

online advertising

social networks

AI and machine learning are less transparent than simple rules

European General Data Protection Regulation (GDPR)

right of access

right to be forgotten

right to explanation