Peter Gasston
C.T. at +rehabstudio
@stopsatgreen
broken-links.com
Slide 3
Slide 3 text
@stopsatgreen
ASCI Red. Intel’s last supercomputer. £43 million (today). Could process 1.8 teraflops. 1,800,000,000,000. Retired in 2000.
Slide 4
Slide 4 text
@stopsatgreen
PS3. Launched in 2005. 1.8 teraflops. £425.
Slide 5
Slide 5 text
@stopsatgreen
In 2012, Google used 16,000 connected processors to teach a computer how to detect a cat in a photo.
>> Jim Wilson, NYT.
Slide 6
Slide 6 text
@stopsatgreen
In 2015, it recognises this as “man using his laptop while his cat looks at the screen”. 5% error rate.
https://twitter.com/timClicks/status/619734363362557953
Slide 7
Slide 7 text
@stopsatgreen
IBM Watson. Beat human contestant at Jeopardy.
10 years of development. Hard to get costs, but at least £12 million.
Slide 8
Slide 8 text
@stopsatgreen
Cognitoys Dino. £80. Powered by Watson.
Slide 9
Slide 9 text
@stopsatgreen
Artificial Intelligence is
becoming very good,
very quickly.
Lots of different strands to AI, but the revolution has been in deep learning.
Slide 10
Slide 10 text
How deep learning works
Artificial neural network.
Data is passed through layers; each layer does a job, then passes the result up to the next layer. Layers can perform pattern analysis or classification.
Can give quick rough answer, slower perfect answer.
Slide 11
Slide 11 text
@stopsatgreen
Computer scientists in Germany have used this to understand the content of an image and the painting styles of old masters.
>> http://arxiv.org/abs/1508.06576
https://github.com/jcjohnson/neural-style
Slide 12
Slide 12 text
@stopsatgreen
B. Turner, C. Van Gogh, D. Munch
Not just adding a filter; it understands the areas of differentiation (sky, buildings, water).
Slide 13
Slide 13 text
@stopsatgreen
Now the AIs are coming
for your job.
Slide 14
Slide 14 text
Is your job at risk?
Abstractions
People
Pictures
Words
Numbers
Routine Variety
You’re alright
Get nervous
Bye Bye
>> http://blogs.gartner.com/martin-kihn/how-to-know-if-a-robot-will-take-your-marketing-job/
Slide 15
Slide 15 text
@stopsatgreen
A shallow magnitude 4.7 earthquake was
reported Monday morning five miles from
Westwood, California, according to the U.S.
Geological Survey. The temblor occurred at
6:25 a.m. Pacific time at a depth of 5.0 miles.
Copywriters are at risk. This is text generated by Quakebot.
Slide 16
Slide 16 text
@stopsatgreen
Kitty couldn’t fall asleep for a long time. Her
nerves were strained as two tight strings, and
even a glass of hot wine, that Vronsky made
her drink, did not help her. Lying in bed she
kept going over and over that monstrous
scene at the meadow.
Russian novel, ‘True Love’, written by computer.
The NYT did a quiz asking you to recognise paras written by robots or humans. I got 2/8 right.
>> http://www.nytimes.com/interactive/2015/03/08/opinion/sunday/algorithm-human-quiz.html?_r=0
@stopsatgreen
One in five of us could be
out of work.
That’s one in five.
Slide 19
Slide 19 text
@stopsatgreen
https://dwnld.me
Slide 20
Slide 20 text
@stopsatgreen
http://www.appmachine.com/instant/
Slide 21
Slide 21 text
@stopsatgreen
https://thegrid.io
Slide 22
Slide 22 text
@stopsatgreen
‘That will just make identikit websites’. Well, people made these. Looks like it could be easily automated.
Slide 23
Slide 23 text
@stopsatgreen
–Travis Gertz
“The work we produce is repeatable and
predictable. It panders to a common
denominator.
We build buckets and templates to hold every
kind of content, then move on to the next
component of the system.
Digital design is a human assembly line.”
Travis Gertz, ‘Design Machines’
https://louderthanten.com/articles/story/design-machines
Slide 24
Slide 24 text
Is your job at risk?
Abstractions
People
Pictures
Words
Numbers
Routine Variety
You’re alright
Get nervous
Bye Bye
A human assembly line…
Slide 25
Slide 25 text
@stopsatgreen
Developers might be feeling a bit smug, but they really shouldn’t be.
Hone lets designers change apps without writing code.
https://hone.tools
Slide 26
Slide 26 text
@stopsatgreen
– Wired
“[muScalpel] successfully transplanted a video
coding format from one media player to
another. It took the automated system 26 hours
to complete the transplant, while VLC's manual
addition of the code happened over a period of
20 days.”
Automated system to take code from one base and move it to another. Currently C only, but could work in any language.
http://www.wired.co.uk/news/archive/2015-07/30/code-organ-transplant-software-myscalpel
Slide 27
Slide 27 text
@stopsatgreen
– MIT news
“Helium is a system that revamps and fine-
tunes code without ever needing the original
source, in a matter of hours or even minutes.”
Performance increases between 75% and 500% in tests.
http://news.mit.edu/2015/computer-program-fixes-old-code-faster-than-expert-engineers-0609
The developers are probably feeling less smug now.
Slide 28
Slide 28 text
@stopsatgreen
AI can’t do our job, but
it can do bits of our job.
We become more efficient, can do more, fewer staff needed. 20% of jobs lost, probably starting with the junior ones.
Slide 29
Slide 29 text
@stopsatgreen
– Travis Gertz
“While we’ve been streamlining our processes
and perfecting our machine-like assembly
techniques, others have been watching closely
and assembling their own machines.
We’ve designed ourselves right into an
environment ripe for automation.”
All of the workflow tools we’ve built make our jobs automatable.
These quotes are from a fairly mandatory article, ‘Design Machines’ by Travis Gertz. But the best thing about the article is that he also suggests a solution to the problem:
Slide 30
Slide 30 text
@stopsatgreen
Slide 31
Slide 31 text
@stopsatgreen
– Travis Gertz
“We need to be better than the machines.
It’s time to step up and design with heart.”
This is the last quote from that article.
Slide 32
Slide 32 text
@stopsatgreen
Deepdream is Google’s tool for visualising neural networks.
https://twitter.com/brendandawes/status/617969546079678464
Slide 33
Slide 33 text
@stopsatgreen
– Cameron Moll
“Mastery of creation and composition is much
more important than mastery of tools.”
http://aneventapart.com/news/post/mastery-of-creation-is-more-important-than-mastery-of-tools
Slide 34
Slide 34 text
@stopsatgreen
– Andrew Ng
“We need to enable a lot of people to do non-
routine, non-repetitive tasks. Teaching
innovation and creativity could be one way to
get there.”
Andrew Ng is a major player in Google and Baidu’s AI. He has major concerns about the effects of automation and AI on employment.
http://www.huffingtonpost.com/2015/05/13/andrew-ng_n_7267682.html
Slide 35
Slide 35 text
@stopsatgreen
We should welcome the
benefits of AI into our builds.
Slide 36
Slide 36 text
@stopsatgreen
– Cennydd Bowles
“AI is becoming a cornerstone of user
experience. This is going to be interesting (read:
difficult) for designers.”
http://www.cennydd.com/blog/ai-and-future-user-experience
Slide 37
Slide 37 text
@stopsatgreen
visual
motion
interaction
experience
service
emotion
design
intelligence
Slide 38
Slide 38 text
@stopsatgreen
– Cennydd Bowles
“We’ll have to create frameworks / scaffolds /
templates for AIs to deliver output through.
These scaffolds may be sonic, tactile, and
linguistic as well as visual.”
Slide 39
Slide 39 text
@stopsatgreen
Much of The Grid’s featureset
is repackaged existing tech.
A lot of what the Grid does is easy (confirm this).
Slide 40
Slide 40 text
@stopsatgreen
Smart Image Cropping.
Slide 41
Slide 41 text
@stopsatgreen
http://thumbor.org
Slide 42
Slide 42 text
@stopsatgreen
Color Thief is good for intelligent colour palettes.
http://lokeshdhakar.com/projects/color-thief/
Slide 43
Slide 43 text
@stopsatgreen
Packery or Susy for intelligent layouts.
http://susy.oddbird.net
http://packery.metafizzy.co
Slide 44
Slide 44 text
@stopsatgreen
We can use AI to provide
better services to our users.
Slide 45
Slide 45 text
@stopsatgreen
Artificial Intelligence is
becoming very good,
very quickly.
Previously the domain of a few internet giants, now many products and services.
Slide 46
Slide 46 text
@stopsatgreen
Artificial Intelligence is
becoming very good,
very quickly.
Artificial Intelligence is
becoming very available,
very quickly.
Previously the domain of a few internet giants, now many products and services.
Slide 47
Slide 47 text
@stopsatgreen
Remember Watson?
Slide 48
Slide 48 text
@stopsatgreen
IBM Developer Cloud gives you access to all of Watson’s capabilities, using Node/Java and a range of RESTFUL APIs.
They just brought a whole range into general availability.
They want to be the ubiquitous platform of AI, as Windows was to the home PC and Android to mobile.
http://www.ibm.com/smarterplanet/us/en/ibmwatson/watson-cloud.html
Slide 49
Slide 49 text
@stopsatgreen
Visual Recognition goes beyond simple face detection, to understand the content of an image. It’s like the Google tool I showed at the start.
Slide 50
Slide 50 text
@stopsatgreen
Has a bizarre fixation on wrestling.
Slide 51
Slide 51 text
@stopsatgreen
El luchador.
Slide 52
Slide 52 text
@stopsatgreen
The more you train
deep learning systems,
the better they get.
Slide 53
Slide 53 text
@stopsatgreen
Google Photos, Flickr, EyeEm all do this.
Slide 54
Slide 54 text
@stopsatgreen
Slide 55
Slide 55 text
@stopsatgreen
At a certain point your users’
expectations will be raised.
Your dumb photo system will seem outdated.
Slide 56
Slide 56 text
@stopsatgreen
New tool released yesterday: visual insights. Give it a batch of images, it will find common themes.
Slide 57
Slide 57 text
@stopsatgreen
Personality Insights find persona characteristics from the content provided by users.
Better targeted promotions.
This is the type of information that other sites gather on you. Now you can gather it on your users!
Slide 58
Slide 58 text
@stopsatgreen
Tradeoff Analytics are used to make comparisons, helping users make better choices between products or objectives.
Slide 59
Slide 59 text
@stopsatgreen
Watson not the only game in town. MS have Project Oxford, the power behind Cortana.
Although you have to know .Net or Android for now.
https://www.projectoxford.ai
Slide 60
Slide 60 text
@stopsatgreen
Has visual recognition. This seems to not be quite as evolved as Watson’s, but also adds ‘racy’ content detection, clever colour extraction, character recognition and
intelligent cropping.
Slide 61
Slide 61 text
@stopsatgreen
Project Oxford also has very good Face APIs, able to detect, recognise, verify, deduce age, find similar faces, and so on.
I like it because it guessed my age as a few years younger than I really am.
Slide 62
Slide 62 text
@stopsatgreen
MetaMind is a new startup with image and text tools.
https://www.metamind.io/
Also http://www.skymind.io/
Slide 63
Slide 63 text
@stopsatgreen
Image recognition seems to get very good results.
Slide 64
Slide 64 text
@stopsatgreen
Natural language processing
is the ‘killer app’ of
deep learning.
Slide 65
Slide 65 text
@stopsatgreen
Natural language understanding to get answers from the content of your site.
Understand what a question is about. Temperature or weather.
Conversations which continue asking until all details are gathered.
Slide 66
Slide 66 text
@stopsatgreen
Alchemy API, an IBM acquisition, analyses text for sentiment, relations, entities… it learns and recommends based on that learning.
Slide 67
Slide 67 text
@stopsatgreen
Pocket and Shutterstock recommendations.
Amazon, Netflix, Google News, etc, all use these types of service.
2/3 of Netflix views are from recommendations.
Slide 68
Slide 68 text
@stopsatgreen
Natural language coming to Spotlight in El Capitan.
Slide 69
Slide 69 text
@stopsatgreen
The real advantage of AI is
in conversation.
@stopsatgreen
This is about more than chatbots.
Chatbots are boring. IKEA have had Anna since at least 2008.
Slide 72
Slide 72 text
@stopsatgreen
–Kyle Dent
“A conversation is a sequence of turns where
each utterance follows from what’s already
been said and is relevant to the overall
interaction.
Dialog systems must maintain a context over
several turns.”
http://radar.oreilly.com/2015/09/talking-to-the-iot.html
Slide 73
Slide 73 text
@stopsatgreen
Virtual assistants are
everybody’s new jam.
Slide 74
Slide 74 text
@stopsatgreen
Amazon Echo is the first home manifestation of the virtual assistant.
Siri in Apple TV, Apple Car, Apple Watch etc.
Google Voice Search, also in Android, Chrome, Wear, Auto, TV, etc.
… in your Windows 10 laptop.
@stopsatgreen
Conversation doesn’t need to
be vocal, but it helps.
Slide 80
Slide 80 text
@stopsatgreen
Making computers talk is easy.
Bell labs, 1961
Slide 81
Slide 81 text
@stopsatgreen
In browsers using the Web Speech API, supported in Chrome, Opera, Safari, Firefox (behind flag)
http://www.broken-links.com/2013/09/20/web-speech-api-part-one-speech-synthesis/
Slide 82
Slide 82 text
@stopsatgreen
Just one of many artificial voice services available.
AT&T, Neospeech.
https://www.ivona.com
http://developer.att.com/apis/speech
http://www.neospeech.com/
Slide 83
Slide 83 text
@stopsatgreen
Making computers listen is
easy… now.
Bell Labs, 1952. Could recognise numbers spoken by one person.
1980s Hidden Markov method.
Recently deep learning.
Slide 84
Slide 84 text
@stopsatgreen
Since 2012 Google’s voice
recognition error rate has
dropped from 26% to 8%.
That’s from one word in 4 to one word in 12. This was a few months ago - a recent update is even better.
Baidu, ‘China’s Google’, says theirs is 6% - one word in 17.
Slide 85
Slide 85 text
@stopsatgreen
10% of Baidu search queries
are by voice.
That’s approx. 500m per day.
Slide 86
Slide 86 text
@stopsatgreen
Web Speech API. Chrome & Firefox OS (soon).
https://developers.google.com/web/updates/2013/01/Voice-Driven-Web-Apps-Introduction-to-the-Web-Speech-API?hl=en
Slide 87
Slide 87 text
@stopsatgreen
Project Oxford native only.
Useful in limited cases.
Slide 88
Slide 88 text
@stopsatgreen
!
All need WebRTC or Flash fallback, so forget Safari/iOS.
Slide 89
Slide 89 text
@stopsatgreen
Making computers understand
is hard. But possible.
Only recognises words (string matching), don’t understand intent.
Slide 90
Slide 90 text
@stopsatgreen
https://www.luis.ai
Slide 91
Slide 91 text
@stopsatgreen
https://api.ai
Slide 92
Slide 92 text
@stopsatgreen
https://wit.ai
Slide 93
Slide 93 text
@stopsatgreen
Slide 94
Slide 94 text
@stopsatgreen
Why natural language is good: train machines, multiple steps, takes minutes for what can be a quite simple query.
Slide 95
Slide 95 text
@stopsatgreen
Even on the web that’s about 10 clicks, some typing - nationalrail.co.uk
Slide 96
Slide 96 text
@stopsatgreen
Note that I don’t have to enter dates manually.
Not all options are available, but I can train it to use them.
I used Chicago because an unfortunate drawback is that they don’t always recognise British English.
Slide 97
Slide 97 text
@stopsatgreen
This is what powers M. AI & HI.
This is the wave of the future.
https://www.facebook.com/Davemarcus/posts/10156070660595195
Slide 98
Slide 98 text
@stopsatgreen
From GUI to CUI.
>> http://www.wired.com/2013/03/conversational-user-interface/
Slide 99
Slide 99 text
@stopsatgreen
http://x.ai
Also https://claralabs.com/
Slide 100
Slide 100 text
@stopsatgreen
There’s a reason that WhatsApp is valued at $19billion.
Shutterstock via http://www.wired.co.uk/news/archive/2014-02/19/whatsapp-exclusive
@stopsatgreen
Conversational UI is an idea
whose time has come.
Natural language understanding is now at the point where this is possible.
Slide 105
Slide 105 text
@stopsatgreen
intelligence
visual
motion
interaction
experience
service
emotion
design
conversation
Slide 106
Slide 106 text
@stopsatgreen
Old problems still apply to
new interaction models.
Slide 107
Slide 107 text
@stopsatgreen
Security is hard. Privacy even harder. Deep learning is impossible without some kind of human training. Samsung TV.
http://www.theguardian.com/technology/2015/feb/27/samsung-voice-recording-smart-tv-breach-privacy-law-campaigners-claim
Slide 108
Slide 108 text
@stopsatgreen
Be transparent.
https://history.google.com/history/audio