Home AGI Papers Scott Aaronson - Why Philosophers Should Care About Computational Complexity (2011)

Scott Aaronson - Why Philosophers Should Care About Computational Complexity (2011)

History / Edit / PDF / EPUB / BIB /
Created: August 24, 2017 / Updated: November 2, 2024 / Status: finished / 15 min read (~2873 words)
Artificial General Intelligence

Complexity theory asks the question: how do the resources needed to solve a problem scale with some measure $n$ of the problem size
- reasonably, like $n$ or $n^2$
- unreasonably, like $2^n$ or $n!$
Theoretical computer scientists generally call an algorithm "efficient" if its running time can be upper-bounded by any polynomial function of $n$, and "inefficient" if its running time can be lower-bounded by any exponential function of $n$

According to the Incompleteness Theorem, one thing that a computer making deductions via fixed formal rules can never do is to "see" the consistency of its own rules. Yet this, they (Lucas and Penrose) assert, is something that human mathematicians can do, via some sort of intuitive perception of Platonic reality
What does one even mean in saying one has a task that "humans can perform but computers cannot"?

In practice, people judge each other to be conscious after interacting for a very short time, perhaps as little as a few seconds
This suggests that we can put a finite upper bound - to be generous, let us say $10^{20}$ - on the number of bits of information that two people A and B would ever realistically exchange, before A had amassed enough evidence to conclude that B was conscious
Given a lookup table of all the possible history H of A and B conversations, we can say that the lookup table would be finite, by the assumption that there is a finite upper bound on the conversation length
This implies that the function $f_B$ is computable
From these simple considerations, we conclude that if there is a fundamental obstacle to computers passing the Turing Test, then it is not to be found in computability theory
(Speculation) (CLAIM1) Any computer program that passed the Turing Test would need to be exponentially-inefficient in the length of the test - as measured in some resource such as time, memory usage, or the number of bits needed to write the program down. In other words, the astronomical lookup table is essentially the best one can do.

An exponential-sized lookup table that passed the Turing Tet would not be sentient (or conscious, intelligent, self-aware, etc.), but a polynomially-bounded program with exactly the same input/output behavior would be sentient
- Furthermore, the latter program would be sentient because it was polynomially-bounded

Are there any reasons to accept the claim (CLAIM1) - the claim that humans are not efficiently simulable by Turing machines?
Despite occasional claims to the contrary, I personally see no reason to believe that humans can solve NP-complete problems in polynomial time, and excellent reasons to believe the opposite
Even if humans were not efficiently simulable by Turing machines, the "direction" in which they were hard to simulate would almost certainly be different from the directions usually considered in complexity theory
Two (hypothetical) ways this could happen
- The task that humans were uniquely good at - like painting or writing poetry - could be incomparable with mathematical tasks like solving NP-complete problems, in the sense that neither was efficiently reducible to the other
- Humans could have the ability to solve interesting special cases of NP-complete problems faster than any Turing machine (e.g., the Pigeonhole Principle, where the Boolean satisfiability algorithms would have to try all variants (exponential))

The problem of logical omniscience: Can we give some formal account of "knowledge" able to accommodate people learning new things without leaving their armchairs?
One vacuous "solution" would be to declare that your knowledge is simply a list of all the true sentences that you "know"
Intuitively, we want to say that your "knowledge" consists of various facts, together with some simple logical consequences of those facts, but not necessarily all the consequences, and certainly not all consequences that involve difficult mathematical reasoning
As soon as we try to formalize this idea, we run into problems
The most obvious problem is the lack of a sharp boundary between the facts you know right away, and those you "could" know, but only after significant thought
A related problem is the lack of a sharp boundary between the facts you know "only if asked about them," and those you know even if you're not asked
There's a third problem that seems much more serious than either of the two above. Namely, you might "know" a particular fact if asked about it one way, but not if asked in a different way
Do you possess an internal algorithm, by which you can answer a large (and possibly-unbounded) set of questions of some form?
- ... answer in a reasonable amount of time ...

Function Polynomial-Time (FP): the set of all functions of natural numbers $f: \mathbb{N} \rightarrow \mathbb{N}$ that are computable in polynomial time by a deterministic Turing machine
FP is "morally" the same as the class P: they differ only in that P is a class of decision problems (or equivalently, functions of $f: \mathbb{N} \rightarrow \{0, 1\}$), whereas FP is a class of functions with integer range
Cobham axioms:
- Every constant function $f$ is efficiently computable, as is every function which is nonzero only finitely often.
- Pairing: If $f(x)$ and $g(x)$ are efficiently computable, then so is $\langle f(x), g(x) \rangle$, where $\langle , \rangle$ is some standard pairing function for the natural numbers.
- Composition: If $f(x)$ and $g(x)$ are efficiently computable, then so is $f(g(x))$.
- Grab Bag: The following functions are all efficiently computable:
  - the arithmetic functions $x + y$ and $x \times y$
  - $|x| = \lfloor \log_2 x \rfloor + 1$ (the number of bits in $x$'s binary representation)
  - the projection functions $\Pi_1 (\langle x, y \rangle) = x$ and $\Pi_2 (\langle x, y \rangle) = y$
  - bit $(\langle x, i \rangle)$ (the $i^{th}$ bit of $x$ binary representation, or 0 if $i > |x|$)
  - diff $(\langle x, i \rangle)$ (the number obtained from $x$ by flipping its $i^{th}$ bit)
  - $2^{|x|^2}$ (called the "smash function")
- Bounded Recursion: Suppose $f(x)$ is efficiently computable, and $|f(x)| \le |x|$ for all $x \in \mathbb{N}$. Then the function $g(\langle x, k \rangle)$, defined by
  $$ g(\langle x, k \rangle) = \left\{ \begin{array}{ll} f(g(\langle x, \lfloor k/2 \rfloor \rangle)) & \mbox{if k > 1} \\ x & \mbox{if k = 1} \end{array} \right.$$
  
  is also efficiently computable.
The class FP, of functions $f: \mathbb{N} \rightarrow \mathbb{N}$ computable in polynomial time by a deterministic Turing machine, satisfies Cobham axioms, and is the smallest class that does so

We could refuse to define an agent's "knowledge" in terms of which individual questions she can quickly answer, and insist of speaking instead about which infinite families of questions she can quickly answer
The main objections to the "infinity cure"
- We've simply pushed the problem of logical omniscience somewhere else
- The infinite families have not been specified
  - The relevant issue will be whether there exists a generalization $G$ of $Q$ (a question), such that the agent knows a fast algorithm for answering questions of type $G$, and also recognize that $Q$ is of type $G$ (e.g., determining whether a number ending in 0 is composite or not, based on the fact that it will be a multiple of 10)
- The relationship between asymptotic complexity and finite statements
  - The usual response: Asymptotic statements are always vulnerable to being rendered irrelevant, if the constant factors turned out to be ridiculous

In this theory, we consider an idealized "learner," who is presented with points $x_1, \dots, x_m$ drawn randomly from some large set $\mathcal{S}$, together with the "classifications" $f(x_1), \dots, f(x_m)$ of those points
The learner's goal is to infer the function $f$, well enough to be able to predict $f(x)$ for most future points $x \in \mathcal{S}$
For simplicity, we often assume that $\mathcal{S}$ is a set of binary strings, and that the function $f$ maps each $x \in \mathcal{S}$ to a single bit, $f(x) \in \{0, 1\}$
The important assumptions are the following:
- Each of the sample points $x_1, \dots, x_m$ is drawn independently from some (possibly-unknown) "sample distribution" $\mathcal{D}$ over $\mathcal{S}$. Furthermore, the future points $x$ on which the learner will need to predict $f(x)$ are drawn from the same distribution
- The function $f$ belongs to a known "hypothesis class" $\mathcal{H}$. This $\mathcal{H}$ represents "the set of possibilities" the learner is willing to entertain" (and is typically much smaller than the set of all $2^{|\mathcal{S}|}$ possible functions from $\mathcal{S}$ to $\{0, 1\}$)
Theorem 2 (Valiant) Consider a finite hypothesis class $\mathcal{H}$, a Boolean function $f : \mathcal{S} \rightarrow \{0, 1\} \in \mathcal{H}$, and a sample distribution $\mathcal{D}$ over $\mathcal{S}$, as well as an error rate $\epsilon > 0$ and a failure probability $\delta > 0$ that the learner is willing to tolerate. Call a hypothesis $h : \mathcal{S} \rightarrow \{0, 1\}$ "good" if
$$ \Pr_{x \sim \mathcal{D}} [h(x) = f(x)] \ge 1 - \epsilon$$
Also, call sample points $x_1, \dots, x_m$ "reliable" if any hypothesis $h \in \mathcal{H}$ that satisfies $h(x_i) = f(x_i)$ for all $i \in \{1, \dots, m\}$ is good. Then
$$ m = \frac{1}{\epsilon}\ln\frac{\mathcal{H}}{\delta}$$
samples points $x_1, \dots, x_m$ drawn independently from $\mathcal{D}$ will be reliable with probability at least $1 - \delta$
In other words, if, by some unspecified means, the learner manages to find any hypothesis $h \in \mathcal{H}$ that makes correct predictions on all its past data points $x_1, \dots, x_m$, then provided $m$ is large enough (and as it happens, $m$ doesn't need to be very large), the learner can be statistically confident that $h$ will also make the correct predictions on most future points

The first drawback is that the Theorem 2 works only for finite hypothesis classes
One could solve this problem by simply discretizing the parameters, but then the number of hypothesis would depend on how fine the discretization was
Fortunately, we can avoid such difficulties by realizing that the learner only cares about the "differences" between two hypothesis insofar as they lead to different predictions
- This leads to the fundamental notion of VC-dimension
Definition 3 (VC-dimension) A hypothesis class $\mathcal{H}$ shatters the sample points ${x_1, \dots, x_k} \subseteq \mathcal{S}$ if for all $2^k$ possible settings of $h(x_1), \dots, h(x_k)$, there exists a hypothesis $h \in \mathcal{H}$ compatible with those settings. Then $VCdim(\mathcal{H})$, the VC-dimension of $\mathcal{H}$, is the largest $k$ for which there exists a subset ${x_1, \dots, x_k} \subseteq \mathcal{S}$ that $\mathcal{H}$ shatters (or if no finite maximum exists, then $VCdim(\mathcal{H}) = \infty$)
Theorem 4 (Blumer et al.) For some universal constant $K > 0$, the bound on $m$ in Theorem 2 can be replaced by
$$ m = \frac{K\ VCdim(\mathcal{H})}{\epsilon}\ln\frac{\mathcal{H}}{\delta\epsilon}$$
with the theorem now holding for any hypothesis class $\mathcal{H}$, finite or infinite
In some sense, Theorem 4 is telling us that finite VC-dimension is a necessary and sufficient condition for scientific induction to be possible
The second drawback of Theorem 2 is that it gives us no clues about how to find a hypothesis $h \in \mathcal{H}$ consistent with the sample data. All it says is that, if we find such an $h$, then $h$ will probably be close to the truth
Learning and cryptography are "dual" problems: a learner wants to find patterns in data, while a cryptographer wants to generate data whose patterns are hard to find
The third drawback of Theorem 2 is the assumption that the distribution $\mathcal{D}$ from which the learner is tested is the same as the distribution from which the sample points were drawn
The goal of science is not merely to summarize observations, and thereby let us make predictions about similar observations. Rather, the goal is to discover explanations with "reach," meaning the ability to predict what would happen in novel or hypothetical situations

When we talk about the simplicity or complexity of hypotheses, we should distinguish two issues:
- The asymptotic scaling of the hypothesis size, as the "size" $n$ of our learning problem goes to infinity
- The constant-factor overheads

Does quantum computing have any implications for philosophy - and specifically, for the interpretation of quantum mechanics?
A scalable quantum computer would test quantum mechanics in an extremely novel regime
The "regime" quantum computers would test is characterized not by an energy scale or a temperature, but by computational complexity
One of the most striking facts about quantum mechanics is that, to represent the state of $n$ entangled particles, one needs a vector of size exponential in $n$

Two possible responses to Detsch's challenge:
- Deny that, if Shor's algorithm works as predicted, that can only be explained by postulating "vast computational resources"
  - Complexity theorists have not yet ruled out the possibility of a fast classical factoring algorithm
- Even if we agree that Shor's algorithm demonstrates the reality of vast computational resources in Nature, it is not obvious that we should think of those resources as "parallel universes"
  - Why not simply say that there is one universe, and that it is quantum-mechanical?

Since the time of Euclid, there have been two main notions of mathematical proof:
- A "proof" is a verbal explanation that induces a sense of certainty (and ideally, understanding) about the statement to be proved, in any human mathematician willing and able to follow it
- A "proof" is a finite sequence of symbols encoding syntactic deductions in some formal system, which start with axioms and end with the statement to be proved
Theoretical computer science deals regularly with a third notion of proof:
- A "proof" is any computational process or protocol (real or imagined) that can terminate in a certain way if and only if the statement to be proved is true

Given two graphs $G$ and $h$, each with $n \approx 10000$ vertices, suppose that an all-powerful but untrustworthy wizard Merlin wishes to convince a skeptical kind Arthur that $G$ and $H$ are not isomorphic
It is not proven that, if $G$ and $H$ are non-isomorphic, there is always a differentiating property that Arthur can verify in time polynomial in $n$
There is something Merlin can do instead: he can let Arthur challenge him
- Arthur, send me a new graph $K$, which you obtained either by randomly permuting the vertices of $G$, or randomly permuting the vertices of $H$. Then I guarantee that I will tell you, without fail, whether $K \cong G$ or $K \cong H$
This protocol has at least four features that merit reflection by anyone interested in the nature of mathematical proofs
- The protocol is probabilistic
- The protocol is interactive
- If Merlin had the power to "peer into Arthur's study" and directly observe whether Arthur started with $G$ or $H$, then clearly he could answer every challenge correctly even if $G \cong H$
- It is zero-knowledge

Multi-prover interactive proofs
Probabilistically checkable proofs
Quantum proofs
Computationally-sound proofs and arguments

Complexity theorists believe that space is more powerful than time
In complexity theory, the difference between space and time manifests itself in the straightforward fact that you can reuse the same memory cells over and over, but you can't reuse the same moments of time

Evolutionary Principle: Knowledge requires a causal process to bring it into existence
NP Hardness Assumption: There is no physical means to solve NP-complete problems in polynomial time

Computational complexity theory can contribute to debates about the foundations of economics by showing that, even in the idealized situation of rational agents who all have perfect information about the state of the world, it will often be computationally intractable for those agents to act in accordance with classical economics

Can we say anything general about when a computational complexity perspective is helpful in philosophy, and when it isn't?
- Computational complexity tends to be helpful when we want to know whether a particular fact does any explanatory work
- Other "philosophical applications" of complexity theory come from the Evolutionary Principle and the NP Hardness Assumption
- Computational complexity tends to be unhelpful when we only want to know whether a particular fact "determines" another fact, and don't care about the length of the inferential chain

Complexity theory only makes asymptotic statements
Many of (what we would like to be) complexity theory's basic principles, such as P != NP, are currently unproved mathematical conjectures, and will probably remain that way for a long time
Complexity theory focuses on only a limited type of computer - the serial, deterministic Turing machine - and fails to incorporate the "messier" computational phenomena found in nature
Complexity theory studies only the worst-case behavior of algorithms, and does not address whether that behavior is representative, or whether it merely reflects a few "pathological" inputs

Aaronson, Scott. "Why philosophers should care about computational complexity." arXiv preprint arXiv:1108.1791 (2011).