A Scientifically-Supportable Notion of Free Will In ... - Scott Aaronson

imminentpoppedAI and Robotics

Feb 23, 2014 (3 years and 5 months ago)

153 views

ADVERTISEMENT

For those who enjoyed the “Memory” session on Monday

Multiplying 10
-
Digit Numbers Using Flickr: The Power
of Recognition Memory

by Andrew Drucker (my PhD student)

http://people.csail.mit.edu/andyd/rec_method.pdf

9883603368




4288997768

42390752785149282624

FREE WILL

Scott Aaronson

Associate Professor
Without Tenure

(!), MIT

The Looniest Talk I’ve Ever Given In My Life

A SCIENTIFICALLY
-
SUPPORTABLE
NOTION OF

IN ONLY
6

CONTROVERSIAL STEPS

Introduction

I’ll present a perspective about free will, quantum
mechanics, and time that I’ve never seen before

Compatibilist? Determinist?
Automaton? No problem! You can
listen to the talk too

I’ll place a much higher premium on being original and
interesting than on being right

Thanks

This talk will assume what David Deutsch calls
the
“momentous dichotomy”
:

Example application:
Quantum computing

Either a given technology is possible, or else there’s
some principled reason why it’s not possible.

Conventional wisdom:

“Free will is a hopelessly muddled
concept. If something isn’t deterministic, then logically, it
must be
random

but a radioactive nucleus obviously
doesn’t have free will!”

But the leap from “indeterminism” to

randomness” here is
total nonsense! In computer science, we deal all the time
with processes that are neither deterministic
nor

random…

Nondeterministic
Finite Automaton

x := x + 5;



// Determinism

x := random(1…10);

// Randomness

x := input();


// “Free will”

Hopelessly
-
Muddled?

Free will

Determinism

We can easily
imagine

“external inputs” to the giant
video game we all live in: the problem is just where such
inputs could
fit

into the actual laws of physics!

Quantum Mechanics and the Brain:

A Bullshit
-
Strewn Interdisciplinary Field

Two obvious difficulties:

(1)
The brain isn’t exactly the most hospitable place for
large
-
scale quantum coherence
(nor is there any clear reason
for such coherence to have evolved)

(2)
Even
if

QM were relevant to brain function, how would
that “help”? Again, randomness


free will

The Deterministic Path of This Talk

1.
A proposed “empirical” notion of free will (based on
algorithmic information theory)

2.
A falsifiable hypothesis about brain function


(Little or no exotic physics needed)

3.
The No
-
Cloning Theorem

4.
Recent applications of the No
-
Cloning Theorem


(Quantum money and copy
-
protected quantum software)

5.

Knightian

uncertainty” about the initial quantum
state of the universe

6.
A radical speculation about time


(Independent motivations from quantum gravity?)

How can we define free will in a way that’s
amenable to scientific investigation?

I propose to consider the question, “Can
machines think?” … The original question,
“Can machines think?” I believe to be too
meaningless to deserve discussion.

A. M. Turing, “Computing Machinery and
Intelligence,”
Mind
, 1950

So Turing immediately replaced it with a
different

question:

“Are there imaginable digital computers
which would do well in the

imitation game
?”

1.

For inspiration, I turned to
computer science’s Prophet

In this talk, I’ll propose a similar “replacement”
for the problem of free will

People mean many different things by “free will”:


-

Legal or moral responsibility


-

The feeling of being in control


-

“Metaphysical freedom”

But arguably, one
necessary condition

for “free will” is
(partial)
unpredictability

not by a hypothetical Laplace
demon, but by actual or conceivable technologies


(DNA testing, brain scanning…)

The Envelope Argument:

If, after you said
anything, you could open a sealed envelope and
read what you just said, that would come pretty
close to an
“empirical refutation of free will”
!

Obviously, many of your actions
are

predictable, and the
fact that they’re predictable doesn’t make them “unfree”!

Discussion

In general, the better someone knows you, the better they
can predict you … but even people who’ve been married
for decades can occasionally surprise each other!
(Otherwise, they would’ve effectively “melded” into a single person)

If someone could predict
ALL

your actions, it seems to me
that you’d be “unmasked as an automaton,” much more
effectively than any philosophical argument could unmask you

But how do we formalize the notion of “predicting
your actions”?

After all, if your actions were perfectly random,
then
in the sense relevant for us
, they’d also be
perfectly predictable!

I’ll solve that problem using a
“Prediction Game”

It’s the year 3000.
You enter the brain
-
scanning machine.

The Prediction Game: Setup Phase

The machine
records all the
neural data it can,
without killing
you

The machine outputs
a self
-
contained
“model” of you

(running on a classical
computer, a quantum
computer, or whatever)

Hardest part of this
whole setup to
formalize!

Q #34: Which physicist would you
least

want to be stranded at sea with: Paul
Davies, Sean Carroll, or Max Tegmark?

The Prediction Game: Testing Phase

“Max Tegmark”

0
0.2
0.4
0.6
0.8
Paul
Sean
Max
Q #35: Multiverse: for or against?

FEEDBACK
LOOP

The Prediction Game: Scoring Phase

The Questions: Q
1
,…,Q
n

Your Answers: A
1
,…,A
n

Predictor’s Guessed Distributions: D
1
,…,D
n

where C = some small constant (like 0.01),

B = the number of bits in the shortest computer
program that outputs A
i

given Q
1
,…,Q
i

and D
1
,…,D
i

as
input, for all i

{1,…,n}

We’ll say the predictor “succeeds” if:

Justification

Beautiful Result from Theory of Algorithmic Randomness
(paraphrase):

Assume you can’t compute anything that’s
Turing
-
uncomputatable. Then the inequality from the last
slide can be satisfied with non
-
negligible probability, in the
limit n

,
if and only if

you’re indeed choosing your
answers randomly according to the predictor’s claimed
distributions D
1
,…,D
n
.

Note:

B is itself an uncomputable quantity! Can
falsify

a
claimed Predictor by computing upper bounds on B, but
never prove absolutely that a Predictor works.

(But the same issue arises for separate reasons, and even arises in QM itself!)

If you don’t like the uncomputable element, can replace B
by the number of bits in the shortest
efficient

program

Crucial Point

In retrospect
, looking back on your entire
sequence of answers A
1
,…,A
n
, the predictor could
always decompose the sequence into (1) a part
that has a small Turing
-
machine description and
(2) a part that’s “algorithmically random.”

But when it’s forced to guess your answers
one by
one
, it might see a third, “fundamentally
unpredictable” component.

So,
can

the Prediction Game be won?

An “aspirational question” that could play a similar
role for neuroscience as the Turing Test plays for AI!

Argument for “yes”:

All information relevant for cognition
seems macroscopic and classical. Even if quantum effects
are present, they should get “washed out as noise”

But this is by no means obvious! Consider the following…

Falsifiable Hypothesis (H):

The behavior of (say) a
mammalian brain, on a ~10s timescale, can be (and
often is) sensitive to molecular
-
level events

2.

If

you believe Hypothesis H, then there would appear to be a
fundamental obstacle to winning the Prediction Game…

“Penrose Lite”:

No speculations here about the brain
as quantum computer, noncomputable QG effects in
microtubules, objective state
-
vector reduction, etc …
just the standard No
-
Cloning Theorem!

3.

The No
-
Cloning Theorem

There’s no general procedure to copy an unknown
quantum state, even approximately

Simple 1
-
Qubit Model Situation

VANILLA

CHOCOLATE

BOXERS

BRIEFS

But can the No
-
Cloning Theorem actually be used to get
quantum states that are both unclonable and
“functional”
?
Recent work in quantum computing theory illustrates that
the answer is yes…

Putting Teeth on the No
-
Cloning Theorem

Quantum Money
(Wiesner 1969, A. 2009,
Farhi et al. 2010, A.
-
Christiano 2011…):
Quantum state |


that a bank can
prepare, people can verify as legitimate,
but counterfeiters can’t copy

Quantum Copy
-
Protected Software
(A. 2009)
: Quantum
state |

f


that a software company can prepare, a customer
can use to compute some function f, but a pirate can’t use to
create more states that also let f be computed

4.

While these proposals raise separate issues (e.g.,
computational complexity), they’re analogous to what we
want in one important respect: if you don’t know how the
state |


or |

f


was prepared, then you can copy it, but
only with
exponentially
-
small success probability

(just like if you were trying to guess the outputs by chance!)

Suppose the Prediction Game can’t be won, even by a being
with unlimited computational power who knows the
dynamical laws of physics (but is constrained by QM).

Then such a being’s knowledge
must

involve Knightian
uncertainty either about the initial state of the universe (say,
at the big bang),
or

about “indexical” questions (e.g., “our”
location within the universe or the Everett multiverse)


For otherwise, the being could win the Prediction Game!

Knightian Uncertainty

5.

In economics,
Knightian uncertainty

means
uncertainty that one can’t even accurately quantify
using probabilities. There are formal tools to
manipulate such uncertainty
(e.g., Dempster
-
Shafer theory)

Poetically, we could think of this
Knightian

uncertainty about initial
conditions as “a place for free will
(or something like it) to hide in a
law
-
governed world”!

“Look, suppose I believed the Prediction Game was
unwinnable. Even so, why would that have
anything

to do
with free will? Even if I don’t know the initial state |

0

,
there still
is

such a state, and combined with the dynamical
laws, it still probabilistically determines the future!”

A Radical Speculation About Time

6.

If the Prediction Game was unwinnable, then it
would seem just as logically coherent to speak about
our decisions determining the initial state, as about
the initial state determining our decisions!

“Backwards
-
in
-
time causation”
, but crucially,
not

of a
sort that can lead to grandfather paradoxes

|0




†††














† †



INITIAL HYPERSURFACE (AT THE BIG BANG?)

MACROSCOPIC
AMPLIFICATION

|

㵼=




䉯戠慳B猠䅬A捥c
潮⁡摡te

|

㵼=




Alice says yes

MACROSCOPIC
AMPLIFICATION

|


|








There’s a “dual description” of the whole spacetime history
that lives on an initial hypersurface only, and that has no
explicit time parameter

just a partially
-
ordered set of
“decisions” about what the quantum state on the initial
hypersurface ought to be.


A decision about particle
A
’s initial state

gets made
“before” a decision about particle
B
’s initial state, if and
only if, in the spacetime history,
A
’s amplification to
macroscopic scale occurs in the causal past of
B
’s
amplification to macroscopic scale

Are there independent reasons, arising from
quantum gravity, to find such a picture attractive?

(Now comes the speculative part of the talk!)

“The Black Hole Free Will Problem”:

You jump into a black
hole. While falling toward the singularity, you decide to wave.

According to
black hole
complementarity
, there’s a “dual
description” living on the event horizon. But how does the event
horizon “know” your decision? Could a
superintelligent

predictor,
by collecting the Hawking radiation, reconstruct your decision
without having ever seen
either

“your” past or “your” future?

The account of free will I’m suggesting can not
only
accommodate

a dual description living one
dimension lower; in some sense, it
demands

such
a description

Two Principles That I Held Inviolate

1.
Evolution from initial to later states is
completely

determined by the Hamiltonian:
there’s no room for free will to “hide” there

2.
Classical memories and records, once written,
can’t be “magically altered” by tinkering with
the universe’s initial state

Without

quantum mechanics
(or some other source
of unclonability)
, my account would have required
abandoning at least one of the principles above!

Conclusions

On the other hand, the idea that the Prediction Game
can

be won
also

strikes me as science fiction!

(For then how could you ever know you were “you,” rather than
one of countless simulations being run by various Predictors?)

I admit: the idea that the Prediction Game can’t be won
(because of, e.g., quantum mechanics and Knightian uncertainty
about the initial state)

strikes me as science fiction

By Deutsch’s “Momentous Dichotomy,”
one of these two
science
-
fiction scenarios has to be right!

Crucially, which scenario is right is not just a metaphysical
conundrum, but something that physics, CS, neurobiology,
and other fields can very plausibly make
progress

on