NOOSCOPE THE RISE OF AI STATISTICAL MODELS AS INSTRUMENTS OF KNOWLEDGEA DIAGRAM OF MACHINE LEARNING ERRORS, BIASES AND LIMITATIONS

VLADAN JOLER AND MATTEO PASQUINELLI (2020)WWW.NOOSCOPE.AI

Naturalisation of bias

Pre-emption fallacyCorrelation as causation

Metadata / Labels

Evaluation bias

Hyperparameters

Multidimensionalvector space

1. Training Dataset

2. Learning AlgorithmModel

3. Model Application

X Statistical Anomaly

X X

X

XX

Underfitting

FittingOverfitting

Accumulated human bias

Dimensionality reduction

HUMAN BIAS, INTERVENTIONS AND ERRORS MACHINE AND STATISTICAL BIAS

In

form

atio

n re

duct

ion

CLASSIFICATION MODALITY GENERATION MODALITY

Undetection of the new

Black box horizon

Generation Input

Gen

erat

ion

Out

put

Past world

Classification Input

Classification Output

Selectionbias

Labelingbias

Historical bias

Representation bias

Hum

an b

ias

Ope

rato

r

Act

ion

Tech

nica

l str

uctu

re

Resolution reduction

Information reduction

Mac

hine

Bia

s

Format framing

Category reduction

Proc

ess

Sensor

Database format

Data

Present world

Operator

Society

Algo

rithm

arc

hite

ctur

e

Topology

Anomaly loss

Approximation

Pre-emption

Stat

istic

al in

fere

nce

Classification Prediction

Pattern recognition

Pattern generation

Accumulated machine bias

Algorithmic statistical Interpolation and Extrapolation

Curve fittingModel fitting

Operator

ExtrapolationInterpolationExtrapolation

Labeling Labeling

Selection Dat

aset

com

posi

tion

Operator

Algorithm

Feature extraction

H

Bi

as a

mpl

ifica

tion

Taxonomies

(Heteromation)

DATA POISONING

DATA ANONIMISATION

Deep dreamingScientific halucination

Future world

(Policeman, scientist, artist)Subject of control

Source selection

Regeneration of the old

Subject of control

(“New Jim Code”)

ADVERSARIAL ATTACK

PRIMER

OBFUSCATION

Ghostworker

OperatorGhostworker

OperatorGhostworker

OperatorGhostworker

Evaluation

Pattern extraction

0 1

Cap

ture

Automation of labour

Calculation of surplus-value

Power of normalisationG

host

Wor

ker

Division of labour

INPU

T

Testing environment