Talk at Open University 28th April 2015
I start with some old work. Regret seems like such a bad emotion, but in unpacking how and why it happens, we find a rich story of finely tuned learning based on the interplay between counterfactual reasoning, raw emotion and low-level learning we share with the simplest creatures. Building this cognitive model a machine learning algorithm improves the rate at which it learns in terms of the number of learning experiences required.
The second part is work in progress, but part way through a similar arc. Theory of mind is usually couched in terms of putting ourselves in an other person's head. However there are good reasons to believe that both ontogenically and phylogenically the awareness of self is in fact an accident of the need to understand other people's models of us. In some sense we see ourselves in the eyes of others. Taking this as at least a plausible model of human consciousness of self, sheds light on potential directions for machine consciousness, emotion and ethics.
3. OU 27 April 2015
University of
Birmingham
Tiree
Tiree Tech Wave
next October 2015
4. OU 27 April 2015
today I am not talking about …
• intelligent internet interfaces
… and dot.com days …
• visualisation and sampling
• situated displays, eCampus,
small device – large display interactions
• fun and games, virtual crackers,
artistic performance, slow time
• creativity and Bad Ideas
• physicality & TouchIT
5. OU 27 April 2015
… or …
Alan
Walks
Wales
learning analytics
flip-classroom
and MOOCs
island data,
heritage and
comms
musicology
and the
long-tail of
small data
6. OU 27 April 2015
… or even lots of lights
http:/www.hcibook.com/alan/projects/firefly/
7. OU 27 April 2015
... but I will talk about
... a few old things about privacy and information
more about ..
understanding regret
the emergence of self work in progress
using computational
modeling
9. OU 27 April 2015
privacy is not monotonic
usual approach – minimise leakage
… but …
restricting / deleting / ignoring some information
may make other information more sensitive
A. J. Dix (1990). Information processing, context and privacy.
Human-Computer Interaction - INTERACT'90.
http://alandix.com/academic/papers/int90/
10. OU 27 April 2015
algorithms have accidental values
machine learning / neural nets may infer rules that:
may not be ethical
and
may not be legal
e.g. jobs and gender discrimination
learning analytics & student progress
A. Dix (1992). Human issues in the use of pattern recognition techniques.
In Neural Networks and Pattern Recognition in Human Computer Interaction.
http://alandix.com/papers/neuro92/neuro92.html
12. OU 27 April 2015
K: “do you know the most destructive force
in the universe”
J: “sugar?”
K: “no regret”
Men in Black 3
13. OU 27 April 2015
why regret?
it seems such a negative emotion
is there some adaptive reason for it?
... or just an accident
14. OU 27 April 2015
features of regret
• modal/counterfactual “what if” analysis
• worst when you ‘nearly’ averted disaster
• seems to be about learning
so how do we learn ....
15. OU 27 April 2015
sensesaction
emotion
(3) evaluation
ow! it hurts!
(4) learnt association
touching thorn
is bad
(1) touch thorn (2) thorn pricks
finger
basic reactions - learning
16. OU 27 April 2015
sensesaction
emotion
(4) veto
(2) learnt association
‘fires’
No action!
(1) about to
touch thorn
(3) bad feeling
basic reactions – moderating action
17. OU 27 April 2015
sensesaction
(3) learnt
association fires
(1) imagination of
planned action
(2) causes similar
brain activity to
actually doing it!
emotion
(4) veto
basic reactions – moderating intention
18. OU 27 April 2015
only works for instant effects
so what about delayed effects?
(e.g. poisonous plant)
need imagination!
19. OU 27 April 2015
sensesaction
emotion
(3) evaluation
“that hurts”
(1) touch plant (2) some time
later your finger
is sore
why?
(4) desire to
make sense
delayed effect – the gap
20. OU 27 April 2015
sensesaction
(7) learnt association
don’t touch that plant
why?
(5) recent salient events
brought to mind
(6) causes simultaneous
activation in
relevant areas
emotion
delayed effect – bringing to mind
21. OU 27 April 2015
sensesaction
(3) evaluation
yuck :-(
(7) learnt association
drinking beer is yucky
(1) drink beer (2) next morning
feel sick
(4) desire to
make sense
why?
(5) recent salient events
brought to mind
(6) causes simultaneous
activation in
relevant areas
emotion
delayed effect – put it together
22. OU 27 April 2015
and now regret ...
similar but also:
causal connections
moderating emotions
23. OU 27 April 2015
sensesaction
emotion
(3) evaluation
yuck :-(
(1) drink beer (2) next morning
feel sick
why?
(4) desire to
make sense
regret – the gap
24. OU 27 April 2015
sensesaction
(7) learnt association
even though action
not obviously linked
or most salient
(5) imagination
causes simultaneous
activation in
relevant areas
emotion
(4) logical deduction of
what mattered
determines what is
brought to mind
(6) causes negative
emotion
“if only I hadn’t”
… regret
regret – casual thinking
25. OU 27 April 2015
sensesaction(7) learnt association
stronger or weaker
depending on
strength of emotion
(5) imagination
causes simultaneous
activation in
relevant areas
emotion
(4) logical deduction of
what mattered
determines what is
brought to mind
(6) logical deduction
of how much
it matters influences
strength of emotion
regret – modifying emotion
26. OU 27 April 2015
but is it true?
if I were a psychologist
I would run an experiment
if I were a brain scientist
I would do a scan
but as a computer scientist ...
... build a computer model
27. OU 27 April 2015
model architecture
game
mechanics
stimulus
cards dealt
response
stick/twist
effect
win/lose
SRE
assoclookup and
choose
emotion
update plug-in
regret
module
post-hoc info.
further cards dealt
modify
basic ML module
28. OU 27 April 2015
it works!
faster (not better) learning
29. OU 27 April 2015
the data
no regret
iteration %best
50 87.47
100 94.43
500 97.27
1000 97.94
with regret
iteration %best
50 90.05
100 97.31
150 97.94
1000 98.60
30. OU 27 April 2015
… and then fixation …
virtual re-exposure …
… and links to dreams, imagination, creativity, etc.
32. OU 27 April 2015
theory of mind
“Little does she know
that I know that she knows
That I know she’s two-timin’ me”
Kursaal Flyers – 1976 (Top 20 hit!)
33. OU 27 April 2015
ToM – conventional view
1 we know our own minds
34. OU 27 April 2015
ToM – conventional view
1 we know our own minds
2 we imagine ourselves
in other’s heads
35. OU 27 April 2015
ToM – conventional view
1 we know our own minds
2 we imagine ourselves
in other’s heads
3 … and attribute thoughts,
intentions, goals
36. OU 27 April 2015
an alternative account
how do we have the cognitive machinery of self?
look for plausible phylogenic process
also ontogenic parallels in child development
37. OU 27 April 2015
predicting actions of animals
1 animals think
and react
2 hunter models
animals’
thoughts
to predict
reactions
38. OU 27 April 2015
predicting actions of humans
1 other people
have models
of us thinking
2 to predict their
reactions we
need model of
them thinking
3 so we get a
model of
ourselves
thinking
39. OU 27 April 2015
self is an accident of sociality
I is in the eye of an other
40. OU 27 April 2015
implications (for AI/robotics) …
consciousness
– model TOM first
– consciousness (of self) may follow
emotion
– interpret others’ emotions & model impact of emotion display
– ‘feelings’ may follow
ethics
– when are machines ethical agents?
(a) when we treat them as such?
(b) when they understand we treat them as such?