How AI can now do more than just copy what humans do
<https://www.rte.ie/author/1049851-marten-kaas/> By Marten Kaas
UCC
Opinion: welcome to a world where artificial intelligence is capable of
learning from experience like we do
Can you teach an old dog new tricks? What about teaching a chess playing
artificial intelligence (AI) to play checkers? Can machines learn and think
in the same way that people do? The answers, at least to the last two
questions, are yes and yes.
Research on AI is not new nor is the idea that we might be able to build an
artificial machine that can think in the same way human beings do. What is
new is that we have recently reached the point where artificially
intelligent agents are capable of learning in the same ways that humans
learn, from experience.
Before the turn of millennium, almost all research in the field of AI
focused on creating AIs that merely imitated human abilities. For example,
although Deep Blue's victory over chess grandmaster Garry Kasparov in 1997
was unprecedented, it was not really a demonstration of a genuine learning
machine.
It was more a demonstration of how a machine with enough specialised
knowledge and computational power could imitate - indeed surpass - the human
ability to learn and master the game of chess. So what changed? Well, we
finally started to design AIs in the way Alan Turing envisioned more than
half a century ago.
Even before the term 'artificial intelligence' was coined Turing made
remarkable insights into the direction research in the field of AI ought to
proceed in pursuit of the creation of thinking machines. His legacy endures
as a result of the thought experiment known as the "Turing Test" but which
Turing called the "Imitation Game.
You might recognise that as the title of the recent movie about Turing,
played by Benedict Cumberbatch, highlighting his role in cracking Enigma,
but which also hinted at Turing's contributions to the creation of the
digital computer.
Insofar as progress on the creation of genuine learning machines is
concerned, most research in the field of AI has been spent attempting to
create machines that imitate human abilities. This is the essence of the
Turing Test: can a human judge, after interacting with two unknown agents,
correctly identify which agent is another human and which is a machine?
Deep Blue is the kind of machine that might be able to pass the Turing Test,
at least as long as the judge observes its chess playing abilities and not
its checkers playing abilities or conversational skills (of which Deep Blue
has none of either).
Research in AI has been focused for too long on simply imitating the adult
human mind. You can see for yourself how far we have yet to go in creating a
chatbot that is capable of passing the Turing Test by talking with Mitsuku
<https://www.pandorabots.com/mitsuku/> .
Four-time winner of the Loebner Prize, awarded to programs considered to be
the most human-like, Mitsuku, though impressive, does not yet convincingly
imitate human conversational abilities.
So despite the fact that the competition for the Loebner Prize was launched
in 1990, this particular competition, indeed most attempts to create AIs
that can pass some version of the Turing Test, do not encourage research
into AI and machine learning per se.
Focusing too intently on the Turing Test has resulted in the erroneous
conflation of imitation with intelligence, the result being the redirection
of resources in the field of AI towards deceptive rather than genuinely
intelligent artificial agents. But we have entered a new age of AI.
Only now, after decades of failed attempts at creating machines that could
pass the Turing Test, can the true genius of Turing's remarks on machine
learning and AI be appreciated.
Turing suggested that we create an artificial agent with the mind of a
child, one that is capable of learning via experience and education. This is
precisely the change that has swept the field of AI over the last two
decades.
The artificial agent AlphaZero developed by DeepMind for example is a
testament to the superhuman heights learning machines can reach if designed
in the way Turing described. AlphaZero learns, tabula rasa, to play chess,
go and shogi (also known as Japanese chess
<http://(https/www.youtube.com/watch?v=7L2sUGcOgh0> ).
Even before AlphaZero, DeepMind developed a general learning agent capable
of playing popular Atari games. The agent learns from the experience it
gains by playing the game as a human would. Watch it learn to play Breakout
<http://(https/www.youtube.com/watch?v=TmPfTpjtdgg> or Space Invaders
<https://www.youtube.com/watch?v=W2CAghUiofY> .
AlphaZero and this new generation of AIs are different in kind from Deep
Blue and the narrow expert systems often associated with AI. Robust
human-like intelligence, as Turing realised, is not achieved through mere
imitation. It is achieved by educating an initially unintelligent artificial
child-mind until it reaches human, or even superhuman, levels of
intelligence.
The hard work that lies before us now is developing a suitable curriculum
for these AIs since, if educated correctly, they could usher in a golden age
the likes of which humanity has never seen before, one that is well worth
striving for.
Can artificial intelligence reduce anxiety in the workplace?
Analysis: tweets allowed researchers to identify participants' personality,
characteristics and anxiety levels
By Souleiman Hasan
<https://www.maynoothuniversity.ie/people/souleiman-hasan> and Jon Gruda
<https://www.maynoothuniversity.ie/people/jon-gruda> , Maynooth University
<http://maynoothuniversity.ie/>
Anxiety levels, particularly at work, can often be overwhelming. It is
widely known that the experience of high anxiety at work can lead to several
negative personal as well as organisational outcomes, including burnout,
sick days, absenteeism and high turnover.
However, the onset of anxiety is often not evident to the person
experiencing it. If we are not able to identify the onset of our own
anxiety, it makes it impossible for us to assess anxiety and the cause of
that anxiety accurately. In addition, if asked to report our anxiety levels
after an important event (such as an important client presentation), we
sometimes forget how anxious we were before. Again, this makes it immensely
difficult to establish a baseline for anxiety over time.
Finally, some people are just more anxious than others, and we need to
consider these individual differences as well.
All these reasons make it difficult to assess anxiety simply by asking
others. In our most recent research
<https://journals.aom.org/doi/10.5465/AMBPP.2018.17625abstract> , we took on
this challenge and developed a machine-learning tool to automatically detect
the onset of anxiety levels through text.
"This research used publicly accessible tweets allow us to measure anxiety
levels in a non-invasive, naturally-occurring setting, and also give us more
insight in the changes over time"
So what does this mean exactly? Well, whenever we write to others about what
we are currently doing, what we are thinking about what or whom we have
talked to, we generate and leave behind clues about our feelings at the
time. For example, when texting friends over the phone, many people do not
use proper punctuation, such as a full stop at the end of every sentence.
However, whenever we're angry or want to express that something is wrong, we
begin to use full stops at the end of sentences, as if to say "pay
attention, I'm using proper punctuation! Something is wrong!"
Although this marks a very small change, and the content of the message does
not change at all, the person they are chatting to immediately knows that
something is wrong. As researchers, we can look back at such texts and make
certain guesses about the texter's feelings at that particular time. Over
time, the accumulation of such clues can be used to identify participants'
personality and characteristics.
Organisations could use an anxiety detection tool to help ensure employees'
well-being and prevent situations of excessive anxiety
An ideal place to look for such clues is Twitter <http://twitter.com/> . It
has over 330 million monthly active users, who produce on average 6,000
tweets per second or 500 million tweets per day. Tweeting is quick, short,
and mostly captures what is going on at any particular moment.
The question then becomes how we can get a machine to score anxiety within
text, so the process can be automated. The answer can be found in the
artificial intelligence domain of machine learning. First, a set of tweets
which are related to the workplace have been distributed to a set of human
raters who uses a scale developed within the psychometrics to measure
anxiety. Secondly, the resulting labelled tweets are used to learn a
software model that can replicate how humans rated the tweets. Finally, the
learned model is run on new tweets and anxiety scores are predicted
This research used publicly accessible tweets to demonstrate the concept as
they allow us to measure anxiety levels in a non-invasive,
naturally-occurring setting, and also give us more insight in the changes of
anxiety in individuals over time. However, microblog-type data can be found
anywhere, and specifically at the workplace. For instance, many
organisations now use Slack <http://slack.com/> , which allows employees to
discuss and share thoughts and expertise on various project issues.
Like with any other technology, cautions have to be taken. Firstly,
artificial intelligence can make errors, which can statistically be more
likely considering an individual piece of text, than with many pieces of
text. Secondly, the privacy of users has to be respected. These two concerns
lead in a direction where data should not be interpreted at an individual
level, but rather in an aggregated anonymised level to find patterns such as
increased anxiety in the organisation overall, for instance.
Organisations could use an anxiety detection tool to help ensure employees'
well-being and prevent situations of excessive anxiety. Knowing that
employees are experiencing increased anxiety levels, organisations could
ensure that training seminars or counselling services are provided to help
reduce anxiety. Such actions could prevent burnout and sick days, as
employees would feel that their employers care about their employees'
well-being.
Dr Souleiman Hasan
<https://www.maynoothuniversity.ie/people/souleiman-hasan> is a Lecturer in
information systems at the School of Business at Maynooth University
<http://maynoothuniversity.ie/> . Dr Jon Gruda
<https://www.maynoothuniversity.ie/people/jon-gruda> is a Lecturer at the
School of Business at Maynooth University <http://maynoothuniversity.ie/>
Why AI will not outsmart us
"Humans may be that fraction of a second slower at translating languages,
recognising speech or describing images, but our reasoning is based on
cause, not association"
<https://www.rte.ie/author/952036-alan-smeaton/> By Alan Smeaton
DCU
Opinion: artificial intelligence will continue to do great things, but only
for limited rather than general tasks
Artificial intelligence systems perform some tasks better than humans.
That's a fact. In our hospitals, for example, AI systems are being used in
medical imaging to analyse images scans to help radiologists to diagnose
tumours that the human eye can miss.
AI is embedded into our national priorities for education and for research.
It's the focus of many of our technology companies and businesses. It's
seeping into advertising, consumer products and into our homes, through
conversational assistants like Amazon
<https://www.amazon.com/Amazon-Echo-Dot-Portable-Bluetooth-Speaker-with-Alex
a-Black/dp/B01DFKC2SO/ref=sr_1_1?ie=UTF8&qid=1531817353&sr=8-1&keywords=Alex
a&smid=ATVPDKIKX0DER> 's Alexa. Almost by stealth, AI is being used to make
decisions for us and about us. What's to stop them making decisions without
us? Could AI systems become so intelligent that they programme themselves to
outsmart us
This concept is referred to as
<http://en.wikipedia.org/wiki/Technological_singularity> "the Singularity",
meaning an uncontrolled, runaway technology growth, a self-fuelling
super-intelligence that takes over the world and changes our civilisation.
This idea that a computer programme could modify itself to do things other
than what it was initially designed for is well-established in software
engineering. So too is the notion of computer software that repairs itself
when it discovers bugs or is infected by a software virus. In theory, we
could have AI singularity, where AI-based programmes improve themselves to
the level of super-intelligence. Some people, like the late Stephen Hawking,
believed
<https://www.theguardian.com/science/2016/oct/19/stephen-hawking-ai-best-or-
worst-thing-for-humanity-cambridge> that this could lead to the end of the
human race.
Let's examine whether it could actually happen. Almost all the AI systems we
currently use are based on matching patterns using very sophisticated
mathematical modelling. Machine learning, deep learning in particular, is no
more than that. These systems use massive amounts of data to detect hidden
patterns. Self-driving cars work because Tesla <http://www.tesla.com/> has
logged over one billion miles of driving data to use as a basis for mining
patterns
Facebook <http://facebook.com/> and Google <http://google.com/> accurately
describe the images and videos we upload because they have billions of our
previous images and videos to mine for patterns of what cats and dogs and
roads and flowers look like. Automated translation across languages works
because systems have been trained to detect the patterns of language use
across different languages. Alexa is able to recognise our speech because
she has been trained to recognise the patterns that exist between sound
waves and the words that are spoken.
The only AI system to approach human intelligence, which is not based on
pattern matching, is IBM <http://www.ibm.com/watson> 's Watson system. This
is a system which reads in and processes large amounts of text information
or even images and then answers questions about what is in the text or
images. It does more than just finding simple facts from the text like names
and dates. It is able to answer complex logical questions that require
inference; the kind that appear in the TV programme Only Connect
<http://en.wikipedia.org/wiki/Only_Connect> . Watson famously beat the world
champion at the US TV game show Jeopardy
<http://en.wikipedia.org/wiki/Jeopardy!> and is now being used in medical
applications, helping to answer questions from doctors about patient
diagnosis based on clinical data in some US hospitals.
While hile Watson's function is very impressive, and accurate, it doesn't
work like the human brain. Watson assumes ALL answers to a question are
possible answers and it calculates a probability for each possible answer by
computing all of these in parallel. Thus, Watson takes a brute force
approach and needs huge computer resources.
Human intelligence may be a bit slower at figuring out answers to questions
as we don't always remember everything quickly enough. We may be that
fraction of a second slower at translating languages, recognising speech or
describing images, but our reasoning is based on cause, not association. We
figure out why things are because we understand causes and effects.
This means that when we understand things like how electricity works or how
the earth revolves around the sun, we understand it based on cause and
effect. We can propel ourselves to higher levels of understanding that are
way more sophisticated than any amount of training data and pattern mining
could reach.
So with this simple but hugely limiting factor, we can be rest assured,
there won't be an AI singularity based on current methodologies. All AI is
doing is very fast pattern-mining for regularities, and intelligence is not
based on the replication of regularities.
AI will continue to do great things, some faster and more reliably than we
can, but only for limited rather than general tasks. Taking over the world
would be a general task category so we're safe - for now.
