1. Jijnasa 2021
AI and the Sense of
Self
Jayati Deshmukh, Srinath Srinivasa
IIIT Bangalore
2. Robot Ethics
Not a futuristic question
Autonomous robotics already part of
our lives
Responsible and ethical behaiour an
imperative
3. Machine
Hermeneutics
Predominant model of inquiry for building
scientific and technological models
Also called the “clockwork” model of the
universe
Universe modeled as an impersonal automaton,
driven by causality
No primacy for elements like free-will, desire,
awareness, intention, etc.
“What we observe is not nature itself, but nature
exposed to our method of inquiry” --Werner
Heisenberg
An automaton of a duck, by Jacques de Vaucanson in 1739.
René Descartes had written, “I have described this earth, and indeed this
whole visible world, as a machine,” and under the spell of the mechanistic
philosophy, even animals and the human body came to be pictured as
machines rather than living organisms.
Image and caption source:
http://www.thesouloftheworld.com/the-new-experim
ent-putting-nature-on-the-rack/
4. Machines vs Societies (Beings)
Made of parts custom built for a specific purpose
Well-defined functionality for each part
Structure designed apriori into its present shape
Imperative design
Made of generic agents capable of playing
several roles
Autonomous actions by agents based on
self-interest and utility maximization
Declarative design
A sense of self
4
5. “Being” Oriented System Design
“Manageable” Complex Systems:
5
Tractable Mangeable Intractable
Machines
Linear
Tractable / Predictable
Dynamics by design
Beings
Resilient Non-Linear
Resilient to initial conditions
Ergodic / Bounded state space
with invariants
Intractable, but manageable due
to invariant stable properties
Chaos
Chaotic Non-Linear
Sensitivity to initial
conditions
Non-Ergodic
Intractable, may have no
invariant / stable states
6. Responsibility and Agency
Prisoners’ Dilemma (PD) represents an abstract
form of conflict between responsibility and
agency.
Responsible choice (C) is strictly dominated by
selfish choice (D) in single-shot transactions
using any solution concept: Nash equilbrium,
Minmax, Iterated dominance.
R: Reward for cooperation
T: Temptation to defect
S: Sucker’s Payoff
P: Penalty for non-cooperation
T > R > P > S
(T+P)/2 > (R+S)/2
7. Paradigms of Responsible AI
Pareto Improvements
Pursue increased returns only as long as it does
not hurt the prospects of other players
Reciprocity / Reflective Equilibrium
“Do unto others as you would have done unto
you” resulting in TIT-FOR-TAT strategy in iterated
PD games. May optionally involve belief revision
resulting in “reflective” reciprocity.
Clearly unfair agreements can also be Pareto
optimal.
8. Paradigms of Responsible AI
Norms and Constraint Satisfaction
Encode responsible outcomes as liveness
constraints that need to be satisfied
Reinforcement Learning
Learn responsible behaviour based on
positive/negative reinforcement signals
Rudimentary RL control loop
Image Source: Google image search
10. Self and Ethics
A sense of ethics and responsibility is
innate in humans and other living
creatures
External reinforcements and social
norms around ethics are a
consequence, rather than a cause of
human ethical behaviour
Innate sense of ethics can be traced to
our elastic sense of self and our
identity set.
Self
Body/mind
complex
(me)
Family
Community
Profession
d=0
d=1
d=2
d=3
11. Computational Transcendence
Sense of self of agent a described as:
S(a): Elastic sense of self of agent a
Ia
: “Identity Set” set of objects that a identifies
with. Ordinarily a ∈ Ia
da
:I → R+
distance function representing semantic
distance to object of identity. Ordinarily d(a,a) = 0
γa
: attenuation factor in the range [0,1] of identity.
Intrinsic “transcendental” utility obtained by
agent a in game state i is given by:
Here, oi
is the extrinsic payoff obtained by object
o ∈ Ia
in game state i
Intuition: a “unit” of self divided across all identity
objects based on attenuation and distance.
14. Transcendence and Emergence of Cooperation
Consider a “transcendental” version of the PD
game, where each player identifies with the
other, with an attenuation level γ
Intrinsic, transcendental utility of any player (say
player A) in game state i is given by:
Given this, we now calculate the expected payoff
for player A on choosing either C or D as follows:
16. Transcendence and the Emergence of Cooperation
Expected (intrinsic) utility of choosing responsible
option (C) overtakes expected (intrinsic) utility of
choosing irresponsible option (D) as
transcendence level increases
At γ = 1, the PD game “flips over”-- making
choosing C just as attractive as choosing D when
γ = 0
Credence to conventional wisdom that enlarging
our sense of self to include the other player,
naturally makes cooperation rational.
Computational Transcendence extended to
other, more complex games and networks with
similar results
CT forms a sound basis for designing an innate
sense of ethics and responsibility in autonomous
AI
17. CT over a 3-player game with collusion
HH HT TH TT
H 6 6 6 6 1 6 1 6 6 10 10 2
T 6 6 1 2 10 10 10 2 10 0 0 0