TopConf Linz, 02/02/2016

29
the friction zone between probability, machine learning and user experience in the consumer IoT @BorisAdryan

Transcript of TopConf Linz, 02/02/2016

the friction zone between probability, machine learning

and user experience in the consumer IoT

@BorisAdryan

MAKEing the Future

http://knowledge.openboxsoftware.com/blog/the-evolution-of-business-intelligence

excerpt from

imagine!

2016: a little reality check

aka

modified, image from http://www.householdappliancesworld.com

health management

air conditioning

smart heating

communications

security

entertainment

lighting controlweather monitoring

room occupancy

health management

air conditioning

smart heating

communications

security

entertainment

lighting controlweather monitoring

room occupancy

individual apps are NOT [ðə ˈfjuːʧə]

source lost, seen on Twitter

we want intelligent thingsthat talk

David Rose: enchanted objects

as UX paradigm

modified, image from http://enchantedobjects.com

a solid scientific

foundation

no magic involved

enchantment has

time

sleep monitor

schedule

location awareness

building control mobilitycapacity

weather

prioritisingplanning

provisioning

acquiring knowledge == learning

machine learning

creative thinking!=

decision makingwithstatistics + algorithms

==

raw datainformation

knowledge

actionable insight

action

reaction

barometric pressure, temperature, coordinates, schedule, …

snow storm coming, airport hotel, need to travel

flying and snow don’t go together

rebook flight

“context”structure

rules

“conversational”dynamicacquired

there’s no absolute truth out there

data

✓hard facts ✓ intuitive

probability

✓ likelihood of some hypothesis being true given the data

30 40 50 60 70average speed at this point [MPH]

time to target [min]

10

20

30

40

50

we have a sense for simple probabilities

who likes more detail?

aka

p-value < 2.2e-16

confidence interval

FDR

posterior probability

95%

0.025

5%

explain this to your neighbour

Battery is going to die today, p < 2.7x10-3

simple x -> y mapping x

k

h

lq

w

g

f-> y mapping

your FitBit temperature

your friend’s dog

the car

computationally, your life is incredibly messy

blog post at https://iot.ghost.io/is-it-all-machine-learning

datatemperature wind speed

wind direction precipitation air pressure airport code

airline aircraft

fully booked? avg delays

cancellations serve booze?

black box

trainingflights

cancelled in the past

classifierranked list of

relevant features

weight of features

thresholds for features

performance metric

new data

prediction

the hypothesis itself is a

mathematical model

explain this to your user

good decisions are based on

experience

machine learning is an iterative

process

training

classifier

performance assessment

good enough?

get on with life

mor

e da

ta fo

r tra

inin

g

data

noyes

from https://hello.is

the issue with missing data

given all features, we can discover

the causality between them

self-learning systems will have to seek ‘missing’ data

other than saying ‘urgent meeting’ in the calendar, how can the system know it’s really urgent?

…preemptively

http://iot.ghost.io

it’s none of your effing business

things getting more creepy…

“Is there something you should tell me, Boris?I thought your wife was travelling…”

…when they’re conversational

life is becoming increasingly dependenton probabilities and abstract quantities

@BorisAdryan

adding to our anxiety of uncertainty,the conversational IoT may potentially feel repetitive, disruptive and intrusive!

quantitative and computational thinking is going to become an essential skill