=Paper=
{{Paper
|id=Vol-2916/paper_5
|storemode=property
|title=On the Differences between Human and Machine Intelligence
|pdfUrl=https://ceur-ws.org/Vol-2916/paper_5.pdf
|volume=Vol-2916
|authors=Roman Yampolskiy
|dblpUrl=https://dblp.org/rec/conf/ijcai/Yampolskiy21a
}}
==On the Differences between Human and Machine Intelligence==
https://ceur-ws.org/Vol-2916/paper_5.pdf
On the Differences between Human and Machine Intelligence
Roman V. Yampolskiy
Computer Science and Engineering, University of Louisville
roman.yampolskiy@louisville.edu
Abstract [Legg and Hutter, 2007a]. However, widespread implicit as-
Terms Artificial General Intelligence (AGI) and Hu- sumption of equivalence between capabilities of AGI and
man-Level Artificial Intelligence (HLAI) have been HLAI appears to be unjustified, as humans are not general
used interchangeably to refer to the Holy Grail of Ar- intelligences. In this paper, we will prove this distinction.
tificial Intelligence (AI) research, creation of a ma- Others use slightly different nomenclature with respect to
chine capable of achieving goals in a wide range of general intelligence, but arrive at similar conclusions. “Lo-
environments. However, widespread implicit assump- cal generalization, or “robustness”: … “adaptation to
tion of equivalence between capabilities of AGI and known unknowns within a single task or well-defined set of
HLAI appears to be unjustified, as humans are not gen- tasks”. … Broad generalization, or “flexibility”: “adapta-
eral intelligences. In this paper, we will prove this dis- tion to unknown unknowns across a broad category of re-
tinction. lated tasks”. …Extreme generalization: human-centric ex-
treme generalization, which is the specific case where the
scope considered is the space of tasks and domains that fit
1 Introduction1
within the human experience. We … refer to “human-cen-
Imagine that tomorrow a prominent technology company tric extreme generalization” as “generality”. Importantly, as
announces that they have successfully created an Artificial we deliberately define generality here by using human cog-
Intelligence (AI) and offers for you to test it out. You decide nition as a reference frame …, it is only “general” in a lim-
to start by testing developed AI for some very basic abilities ited sense. … To this list, we could, theoretically, add one
such as multiplying 317 by 913, and memorizing your phone more entry: “universality”, which would extend “general-
number. To your surprise, the system fails on both tasks. ity” beyond the scope of task domains relevant to humans,
When you question the system’s creators, you are told that to any task that could be practically tackled within our uni-
their AI is human-level artificial intelligence (HLAI) and as verse (note that this is different from “any task at all” as un-
most people cannot perform those tasks neither can their AI. derstood in the assumptions of the No Free Lunch theorem
In fact, you are told, many people can’t even compute 13 x [Wolpert and Macready, 1997; Wolpert, 2012]).” [Chollet,
17, or remember name of a person they just met, or recog- 2019].
nize their coworker outside of the office, or name what they
had for breakfast last Tuesday2. The list of such limitations 2 Prior work
is quite significant and is the subject of study in the field of
Artificial Stupidity [Trazzi and Yampolskiy, 2018; Trazzi We call some problems ‘easy’, because they come naturally
and Yampolskiy, 2020]. to us like understanding speech or walking and we call other
Terms Artificial General Intelligence (AGI) [Goertzel et problems ‘hard’ like playing Go or violin, because those are
al., 2015] and Human-Level Artificial Intelligence (HLAI) not human universals and require a lot of talent and effort
[Baum et al., 2011] have been used interchangeably (see [Yampolskiy, 2012]. We ignore ‘impossible’ for humans to
[Barrat, 2013], or “(AGI) is the hypothetical intelligence of master domains, since we mostly don’t even know about
a machine that has the capacity to understand or learn any them or see them as important. As LeCun puts it: “[W]e can't
intellectual task that a human being can.” [Anonymous, imagine tasks that are outside of our comprehension, right,
Retrieved July 3, 2020]) to refer to the Holy Grail of Artifi- so we think, we think we are general, because we're general
cial Intelligence (AI) research, creation of a machine capa- of all the things that we can apprehend, but there is a huge
ble of: achieving goals in a wide range of environments world out there of things that we have no idea” [LeCun,
August 31, 2019]. Others, agree: “we might not even be
2
1
Copyright © 2021 for this paper by its authors. Use permitted Some people could do that and more, for example 100,000 digits
under Creative Commons License Attribution 4.0 International of π have been memorized using special mnemonics.
(CC BY 4.0).
aware of the type of cognitive abilities we score poorly on.” such a thing as universal intelligence, and if human intelli-
[Barnett, December 23, 2019]. gence is an implementation of it, then this algorithm of uni-
This is most obvious in how we test for intelligence. For versal intelligence should be the end goal of our field, and
example, Turing Test [Turing, 1950], by definition, doesn’t reverse-engineering the human brain could be the shortest
test for universal general intelligence, only for human-level path to reach it. It would make our field close-ended: a riddle
intelligence in human domains of expertise. Like a drunkard to be solved. If, on the other hand, human intelligence is a
searching for his keys under the light because there it is eas- broad but ad-hoc cognitive ability that generalizes to hu-
ier to find them, we fall for the Streetlight effect observation man-relevant tasks but not much else, this implies that AI is
bias only searching for intelligence in domains we can easily an open-ended, fundamentally anthropocentric pursuit, tied
comprehend [Yampolskiy, 2019]. “The g factor, by defini- to a specific scope of applicability.” [Chollet, 2019].
tion, represents the single cognitive ability common to suc- Humans have general capability only in those human ac-
cess across all intelligence tests, emerging from applying cessible domains and likewise artificial neural networks in-
factor analysis to test results across a diversity of tests and spired by human brain architecture do unreasonably well in
individuals. But intelligence tests, by construction, only en- the same domains. Recent work by Tegmark et al. shows
compass tasks that humans can perform – tasks that are im- that deep neural networks would not perform as well in ran-
mediately recognizable and understandable by humans (an- domly generated domains as they do in those domains hu-
thropocentric bias), since including tasks that humans mans consider important, as they map well to physical prop-
couldn’t perform would be pointless. Further, psychomet- erties of our universe. “We have shown that the success of
rics establishes measurement validity by demonstrating pre- deep and cheap (low-parameter-count) learning depends not
dictiveness with regard to activities that humans value (e.g. only on mathematics but also on physics, which favors cer-
scholastic success): the very idea of a “valid” measure of tain classes of exceptionally simple probability distributions
intelligence only makes sense within the frame of reference that deep learning is uniquely suited to model. We argued
of human values.” [Chollet, 2019]. that the success of shallow neural networks hinges on sym-
Moravec further elaborates the difference between future metry, locality, and polynomial log-probability in data from
machines and humans: “Computers are universal machines, or inspired by the natural world, which favors sparse low-
their potential extends uniformly over a boundless expanse order polynomial Hamiltonians that can be efficiently ap-
of tasks. Human potentials, on the other hand, are strong in proximated.” [Lin et al., 2017].
areas long important for survival, but weak in things far re-
moved. Imagine a “landscape of human competence,” hav- 3 Humans are not AGI
ing lowlands with labels like “arithmetic” and “rote memo-
An agent is general (universal [Hutter, 2004]) if it can learn
rization,” foothills like “theorem proving” and “chess play-
anything another agent can learn. We can think of a true AGI
ing,” and high mountain peaks labeled “locomotion,”
agent as a superset of all possible NAIs (including capacity
“hand-eye coordination” and “social interaction.” Advanc-
to solve AI-Complete problems [Yampolskiy, 2013]). Some
ing computer performance is like water slowly flooding the
agents have limited domain generality, meaning they are
landscape. A half century ago it began to drown the low-
general, but not in all possible domains. The number of do-
lands, driving out human calculators and record clerks, but
mains in which they are general may still be Dedekind-infi-
leaving most of us dry. Now the flood has reached the foot-
nite, but it is a strict subset of domains in which AGI is ca-
hills, and our outposts there are contemplating retreat. We
pable of learning. For an AGI it’s domain of performance is
feel safe on our peaks, but, at the present rate, those too will
any efficiently learnable capability, while humans have a
be submerged within another half century.” [Moravec,
smaller subset of competence. Non-human animals in turn
1998].
may have an even smaller repertoire of capabilities, but are
Chollet writes: “How general is human intelligence? The
nonetheless general in that subset. This means that humans
No Free Lunch theorem [Wolpert and Macready, 1997;
can do things animals cannot and AGI will be able to do
Wolpert, 2012] teaches us that any two optimization algo-
something no human can. If an AGI is restricted only to do-
rithms (including human intelligence) are equivalent when
mains and capacity of human expertise, it is the same as
their performance is averaged across every possible prob-
HLAI.
lem, i.e. algorithms should be tailored to their target problem
Humans are also not all in the same set, as some are ca-
in order to achieve better-than-random performance. How-
pable of greater generality (G factor [Jensen, 1998]) and can
ever, what is meant in this context by “every possible prob-
succeed in domains, in which others cannot. For example,
lem” refers to a uniform distribution over problem space; the
only a tiny subset of all people is able to conduct cutting-
distribution of tasks that would be practically relevant to our
edge research in quantum physics, implying differences in
universe (which, due to its choice of laws of physics, is a
our general capabilities between theory and practice. While
specialized environment) would not fit this definition. Thus
theoretical definition of general intelligence is easy to un-
we may ask: is the human g factor universal? Would it gen-
derstand, its practical implementation remains uncertain.
eralize to every possible task in the universe? … [T]his
question is highly relevant when it comes to AI: if there is
“LeCun argues that even self-supervised learning and learn- fibers. So now what comes to your, to your brain, is a fixed
ings from neurobiology won’t be enough to achieve artifi- but random permutation of all the pixels, there's no way in
cial general intelligence (AGI), or the hypothetical intelli- hell that your visual cortex, even if I do this to you in in-
gence of a machine with the capacity to understand or learn fancy, will actually learn vision to the same level of quality
from any task. That’s because intelligence — even human that you can.” [LeCun, August 31, 2019].
intelligence — is very specialized, he says. “AGI does not Chollet elaborates on the subject of human unlearnable
exist — there is no such thing as general intelligence,” said tasks: “[H]uman intellect is not adapted for the large major-
LeCun. “We can talk about rat-level intelligence, cat-level ity of conceivable tasks. This includes obvious categories of
intelligence, dog-level intelligence, or human-level intelli- problems such as those requiring long-term planning be-
gence, but not artificial general intelligence.”” [Wiggers, yond a few years, or requiring large working memory (e.g.
May 2, 2020]. multiplying 10-digit numbers). This also includes problems
An agent is not an AGI equivalent if it could not learn for which our innate cognitive priors are unadapted; … For
something another agent could learn. Hence, we can divide instance, in the [Traveling Salesperson Problem] TSP, hu-
all possible tasks into human learnable and those, which no man performance degrades severely when inverting the goal
human can learn, establishing that humans are not AGI from “finding the shortest path” to “finding the longest path”
equivalent. We already described ‘easy’ and ‘hard’ for hu- [MacGregor and Ormerod, 1996] – humans perform even
mans problems, the third category of ‘impossible’ is what worse in this case than one of the simplest possible heuristic:
we would classify as abilities impossible for humans to learn farthest neighbor construction. A particularly marked hu-
efficiently [Valiant, 2013]. Computer-unaided humans man bias is dimensional bias: humans … are effectively un-
[Blum and Vempala, 2020] do not possess capabilities in able to handle 4D and higher. … Thus, … “general intelli-
this category, to any degree, and are unlikely to be able to gence” is not a binary property which a system either pos-
learn them. If performed by a human, they would be consid- sesses or lacks. It is a spectrum,” [Chollet, 2019]. “Human
ered magical, but as Arthur Clarke has famously stated: physical capabilities can thus be said to be “general”, but
“Any sufficiently advanced technology is indistinguishable only in a limited sense; when taking a broader view, humans
from magic.” reveal themselves to be extremely specialized, which is to
Some current examples include: estimating face from be expected given the process through which they evolved.”
speech [Oh et al., 2019], DNA [Sero et al., 2019] or ear [Chollet, 2019]. “[W]e are born with priors about ourselves,
[Yaman et al., 2020], extracting passwords from typing about the world, and about how to learn, which determine
sounds [Zhuang et al., 2009; Shumailov et al., 2019], using what categories of skills we can acquire and what categories
lightbulbs [Nassi et al., 2020] and hard drives [Kwong et al., of problems we can solve.” [Chollet, 2019].
2019] as microphones, communicating via heat emissions If such tasks are in fact impossible for any human to per-
[Guri et al., 2015b], or memory-write-generated electro- form, that proves that humans are not AGI equivalent. But,
magnetic signals [Guri et al., 2015a], and predicting gender, how do we know what a highly intelligent agent is capable
age and smoking status from images of retinal fundus of or more interestingly incapable of learning? How do we
[Poplin et al., 2018]. This is what is already possible with know what human’s can’t learn [Ziesche and Yampolskiy,
Narrow AI (NAI) today, AGI will be able to see patterns 2020]? One trick we can use, is to estimate the processing
where humans see nothing but noise, invent technologies we speed [Roberts and Stankov, 1999] for an average human on
never considered possible and discover laws of physics far a particular learning task and to show that even 120 years, a
above our understanding. Capabilities, we humans will very optimistic longevity estimate for people, is not suffi-
never possess, because we are not general intelligences. cient to complete learning that particular task, while much
Even humans armed with simple calculators are no match faster computer can do so in seconds.
for such problems. Generality can be domain limited or unlimited. Different
LeCun gives an example of one task no human could animals, such as dolphins, elephants, mice, etc. and humans
learn: “So let me take a very specific example, it's not an are all general in overlapping but not identical sets of do-
example it's more like a quasi-mathematical demonstration, mains. Humans are not a superset of all animal intelligences.
so you have about 1 million fibers coming out of one of your There are some things animals can do that humans cannot
eyes, okay two million total, but let's talk about just one of and vice versa. For example, humans can’t learn to speak
them. It's 1 million nerve fibers in your optical nerve, let's animal “languages” and animals can’t learn to play chess
imagine that they are binary so they can be active or inac- [Yampolskiy, 2018b]. Richard Hamming made this point in
tive, so the input to your visual cortex is 1 million bits. Now, his famous paper - “The Unreasonable Effectiveness of
they connected to your brain in a particular way and your Mathematics”: "Just as there are odors that dogs can smell
brain has connections that are kind of a little bit like a con- and we cannot, as well as sounds that dogs can hear and we
volution net they are kind of local, you know, in the space cannot, so too there are wavelengths of light we cannot see
and things like this. Now imagine I play a trick on you, it's and flavors we cannot taste. Why then, given our brains
a pretty nasty trick I admit, I cut your optical nerve and I put wired the way they are, does the remark, "Perhaps there are
a device that makes a random permutation of all the nerve thoughts we cannot think," surprise you? Evolution, so far,
may possibly have blocked us from being able to think in as a single agent. This may have been true for a number of
some directions; there could be unthinkable thoughts." years now, and is becoming more definitive every year. As
[Wigner, 1990]. an AI agent can be a superset of many algorithms from
Only AGI is universal/general intelligence over all learn- which it can choose it would not be a subject to the No Free
able domains. AGI is not just capable of anything a human Lunch (NFL) theorems [Wolpert and Macready, 1997;
can do; it is capable of learning anything that could be Wolpert, 2012].
learned. It is a Superset of all NAIs and is equal in capability While AI dominates humans in most domains of human
to Superintelligence. interest [Goodfellow et al., 2014; Mnih et al., 2015; Silver
et al., 2017; Devlin et al., 2018; Clark et al., 2019; Vinyals
4 Conclusions et al., 2019], there are domains in which humans would not
even be able to meaningfully participate. This is similar to
There is no shortage of definitions of intelligence [Legg and
the Unpredictability [Yampolskiy, 2020b] and Unexplaina-
Hutter, 2007a; Legg and Hutter, 2007b; Hernández-Orallo,
bility/Incomprehensibility of AI [Yampolskiy, 2019] re-
2017; Wang, 2019; Yampolskiy, 2020a], but we felt it was
sults, but at a meta-level. The implications for AI control
important to clarify that humans are neither fully general nor
and AI Safety and Security [Callaghan et al., 2017;
terminal point in the space of the possible minds
Yampolskiy, 2018a; Babcock et al., 2019; Babcock et al.,
[Yampolskiy, 2015]. As Chollet says: “We may even build
July 16-19, 2016] are not encouraging. To be dangerous AI
systems with higher generalization power (as there is no a
doesn’t have to be general, it is sufficient for it to be superior
priori reason to assume human cognitive efficiency is an up-
to humans in a few strategic domains. If AI can learn a par-
per bound), or systems with a broader scope of application.
ticular domain it will quickly go from Hypohuman to Hy-
Such systems would feature intelligence beyond that of hu-
perhuman performance [Hall, 2009]. Additionally, common
mans.” [Chollet, 2019]. Humans only have a subset of capa-
proposal for merging of humanity with machines doesn’t
bilities an AGI will have and the capability difference be-
seem to work as adding HLAI to AGI adds nothing to AGI,
tween us and AGI is far greater than capability difference
meaning in a cyborg agent human will become a useless bot-
between AGI and superintelligence (SAI). Bostrom de-
tleneck as AI becomes more advanced and the human will
scribes three forms of superintelligence (p. 53-57)
be eventually removed, if not explicitly at least implicitly
[Bostrom, 2014]: Speed SAI (like a faster human), Collec-
from control. What does this paper tell us? Like the dark
tive SAI (like a group of humans), and Quality SAI (does
matter of the physical universe, the space of all problems is
what humans can’t). All three can be accomplished by an
mostly unknown unknowns, and most people don’t know
AGI, so there is no difference between AGI and SAI, they
that and don’t even know that they don’t know it. To para-
are the same (HLAI ≤ AGI = SAI) and the common takeoff-
phrase the famous saying: “The more AI learns, the more I
speed debate [Yudkowsky and Hanson, 2008] resolves to
realize how much I don't know.”
hard takeoff, from definitions. This implies even stronger
limitations [Yampolskiy, 2017; Yampolskiy, 2019;
Yampolskiy, 2020b] on our capability to control AI and a References
more immediate faceoff. We are already having many prob-
lems with Ignorance Explosion [Lukasiewicz, 1974; [Anonymous, Retrieved July 3, 2020] Anonymous. Artifi-
Lukasiewicz, 1994], an Intelligence Explosion [Loosemore cial general intelligence. Wikipedia. Available at:
https://en.wikipedia.org/wiki/Artificial_general_intelli-
and Goertzel, 2012; Muehlhauser and Salamon, 2012] will gence.
be well beyond our capabilities to control.
[Babcock et al., July 16-19, 2016] James Babcock, Janos
If we use Legg’s definition of intelligence [Legg and Kramar and Roman Yampolskiy. The AGI Containment
Hutter, 2007a], and average performance across all possible Problem. The Ninth Conference on Artificial General In-
problems, we can arrive at a somewhat controversial result telligence (AGI2015). NYC, USA.
that modern AI is already smarter than any human is. An [Babcock et al., 2019] James Babcock, János Kramár and
individual human can only learn a small subset of domains Roman V Yampolskiy. Guidelines for artificial intelli-
and human capabilities can’t be trivially transferred between gence containment. Next-Generation Ethics: Engineer-
different humans to create a union function of all human ca- ing a Better Society. A. E. Abbas: 90-112.
pabilities, but that is, at least theoretically, possible for AI. [Barnett, December 23, 2019] Matthew Barnett. Might hu-
Likewise, humans can’t emulate some computer algorithms, mans not be the most intelligent animals? Available at:
but computers can run any algorithm a human is using. Ma- https://www.less-
chines of 2020 can translate between hundreds of languages, wrong.com/posts/XjuT9vgBfwXPxsdfN/might-hu-
mans-not-be-the-most-intelligent-animals.
win most games, generate art, write poetry and learn many
tasks individual humans are not capable of learning. If we [Barrat, 2013] James Barrat. Our final invention: Artificial
intelligence and the end of the human era, Macmillan.
were to integrate all such abilities into a single AI agent it
would on average outperform any person across all possible [Baum et al., 2011] Seth D Baum, Ben Goertzel and Ted G
Goertzel. "How long until human-level AI? Results from
problem domains, but perhaps not humanity as a whole seen
an expert assessment." Technological Forecasting and [Legg and Hutter, 2007a] Shane Legg and Marcus Hutter.
Social Change 78(1): 185-195. "A collection of definitions of intelligence." Frontiers in
[Blum and Vempala, 2020] Manuel Blum and Santosh Artificial Intelligence and applications 157: 17.
Vempala. "The complexity of human computation via a [Legg and Hutter, 2007b] Shane Legg and Marcus Hutter.
concrete model with an application to passwords." Pro- "Universal intelligence: A definition of machine intelli-
ceedings of the National Academy of Sciences 117(17): gence." Minds and Machines 17(4): 391-444.
9208-9215. [Lin et al., 2017] Henry W Lin, Max Tegmark and David
[Bostrom, 2014] Nick Bostrom. Superintelligence: Paths, Rolnick. "Why does deep and cheap learning work so
dangers, strategies, Oxford University Press. well?" Journal of Statistical Physics 168(6): 1223-1247.
[Callaghan et al., 2017] Vic Callaghan, James Miller, Ro- [Loosemore and Goertzel, 2012] Richard Loosemore and
man Yampolskiy and Stuart Armstrong. The Technolog- Ben Goertzel. Why an intelligence explosion is probable.
ical Singularity: Managing the Journey, Springer. Singularity hypotheses, Springer: 83-98.
[Chollet, 2019] François Chollet. "On the measure of intel- [Lukasiewicz, 1974] Julius Lukasiewicz. "The ignorance
ligence." arXiv preprint arXiv:1911.01547. explosion." Leonardo 7(2): 159-163.
[Clark et al., 2019] Peter Clark, Oren Etzioni, Tushar Khot, [Lukasiewicz, 1994] Julius Lukasiewicz. The ignorance ex-
Bhavana Dalvi Mishra, Kyle Richardson, Ashish Sa- plosion: Understanding industrial civilization, McGill-
bharwal, Carissa Schoenick, Oyvind Tafjord, Niket Tan- Queen's Press-MQUP.
don and Sumithra Bhakthavatsalam. "From 'F' to 'A' on [MacGregor and Ormerod, 1996] James N MacGregor and
the NY Regents Science Exams: An Overview of the Tom Ormerod. "Human performance on the traveling
Aristo Project." arXiv preprint arXiv:1909.01958. salesman problem." Perception & psychophysics 58(4):
[Devlin et al., 2018] Jacob Devlin, Ming-Wei Chang, Ken- 527-539.
ton Lee and Kristina Toutanova. "Bert: Pre-training of [Mnih et al., 2015] Volodymyr Mnih, Koray Kavukcuoglu,
deep bidirectional transformers for language understand- David Silver, Andrei A Rusu, Joel Veness, Marc G Bel-
ing." arXiv preprint arXiv:1810.04805. lemare, Alex Graves, Martin Riedmiller, Andreas K
[Goertzel et al., 2015] Ben Goertzel, Laurent Orseau and Ja- Fidjeland and Georg Ostrovski. "Human-level control
vier Snaider. "Artificial General Intelligence." Scholar- through deep reinforcement learning." Nature
pedia 10(11): 31847. 518(7540): 529-533.
[Goodfellow et al., 2014] Ian Goodfellow, Jean Pouget-Ab- [Moravec, 1998] Hans Moravec. "When will computer
adie, Mehdi Mirza, Bing Xu, David Warde-Farley, hardware match the human brain." Journal of evolution
Sherjil Ozair, Aaron Courville and Yoshua Bengio. Gen- and technology 1(1): 10.
erative adversarial nets. Advances in neural information [Muehlhauser and Salamon, 2012] Luke Muehlhauser and
processing systems. Anna Salamon. Intelligence explosion: Evidence and im-
[Guri et al., 2015a] Mordechai Guri, Assaf Kachlon, Ofer port. Singularity hypotheses, Springer: 15-42.
Hasson, Gabi Kedma, Yisroel Mirsky and Yuval Elovici. [Nassi et al., 2020] Ben Nassi, Yaron Pirutin, Adi Shamir,
GSMem: Data Exfiltration from Air-Gapped Computers Yuval Elovici and Boris Zadov. Lamphone: Real-Time
over GSM Frequencies. 24th USENIX Security Sympo- Passive Sound Recovery from Light Bulb Vibrations.
sium (USENIX Security 15). Cryptology ePrint Archive. Available at:
[Guri et al.,2015b] Mordechai Guri, Matan Monitz, Yisroel https://eprint.iacr.org/2020/708.
Mirski and Yuval Elovici. Bitwhisper: Covert signaling [Oh et al., 2019] Tae-Hyun Oh, Tali Dekel, Changil Kim,
channel between air-gapped computers using thermal Inbar Mosseri, William T Freeman, Michael Rubinstein
manipulations. 2015 IEEE 28th Computer Security and Wojciech Matusik. Speech2face: Learning the face
Foundations Symposium, IEEE. behind a voice. Proceedings of the IEEE Conference on
[Hall, 2009] J Storrs Hall. Beyond AI: Creating the con- Computer Vision and Pattern Recognition.
science of the machine, Prometheus books. [Poplin et al., 2018] Ryan Poplin, Avinash V. Varadarajan,
[Hernández-Orallo, 2017] José Hernández-Orallo. The Katy Blumer, Yun Liu, Michael V. McConnell, Greg S.
measure of all minds: evaluating natural and artificial in- Corrado, Lily Peng and Dale R. Webster. "Prediction of
telligence, Cambridge University Press. cardiovascular risk factors from retinal fundus photo-
[Hutter, 2004] Marcus Hutter. Universal artificial intelli- graphs via deep learning." Nature Biomedical Engineer-
gence: Sequential decisions based on algorithmic proba- ing 2(3): 158-164.
bility, Springer Science & Business Media. [Roberts and Stankov, 1999] Richard D Roberts and Lazar
[Jensen, 1998] Arthur Robert Jensen. The g factor: The sci- Stankov. "Individual differences in speed of mental pro-
ence of mental ability, Praeger Westport, CT. cessing and human cognitive abilities: Toward a taxo-
nomic model." Learning and Individual Differences
[Andrew et al., 2019] Andrew Kwong, Wenyuan Xu and 11(1): 1-120.
Kevin Fu. Hard drive of hearing: Disks that eavesdrop
with a synthesized microphone. 2019 IEEE Symposium [Sero et al., 2019] Dzemila Sero, Arslan Zaidi, Jiarui Li, Ju-
on Security and Privacy (SP), IEEE. lie D White, Tomás B González Zarzar, Mary L
Marazita, Seth M Weinberg, Paul Suetens, Dirk Vander-
[LeCun, August 31, 2019] Yann LeCun. Yann LeCun: Deep meulen and Jennifer K Wagner. "Facial recognition from
Learning, Convolutional Neural Networks, and Self-Su- DNA using face-to-DNA classifiers." Nature communi-
pervised Learning. AI Podcast, Lex Fridman. Available cations 10(1): 2557.
at: https://www.youtube.com/watch?v=SGSOCuByo24.
[Shumailov et al., 2019] Ilia Shumailov, Laurent Simon, [Yampolskiy, 2018a] Roman V Yampolskiy. Artificial
Jeff Yan and Ross Anderson. "Hearing your touch: A Intelligence Safety and Security, Chapman and
new acoustic side channel on smartphones." arXiv pre- Hall/CRC.
print arXiv:1903.11137. [Yampolskiy, 2018b] Roman V Yampolskiy. "The
[Silver et al., 2017] David Silver, Julian Schrittwieser, Ka- singularity may be near." Information 9(8): 190.
ren Simonyan, Ioannis Antonoglou, Aja Huang, Arthur [Yampolskiy, 2019] Roman V Yampolskiy.
Guez, Thomas Hubert, Lucas Baker, Matthew Lai and "Unexplainability and incomprehensibility of artificial
Adrian Bolton. "Mastering the game of go without hu- intelligence." arXiv preprint arXiv:1907.03869.
man knowledge." Nature 550(7676): 354.
[Yampolskiy, 2020a] Roman V Yampolskiy. "On Defining
[Trazzi and Yampolskiy, 2018] Michaël Trazzi and Roman Differences Between Intelligence and Artificial
V Yampolskiy. "Building safer AGI by introducing arti- Intelligence." Journal of Artificial General Intelligence
ficial stupidity." arXiv preprint arXiv:1808.03644. 11(2): 68-70.
[Trazzi and Yampolskiy, 2020] Michaël Trazzi and Roman [Yampolskiy, 2020b] Roman V Yampolskiy.
V Yampolskiy. "Artificial Stupidity: Data We Need to "Unpredictability of AI: On the Impossibility of
Make Machines Our Equals." Patterns 1(2): 100021. Accurately Predicting All Actions of a Smarter Agent."
[Turing, 1950] A. Turing. "Computing Machinery and Intel- Journal of Artificial Intelligence and Consciousness
ligence." Mind 59(236): 433-460. 7(01): 109-118.
[Valiant, 2013] Leslie Valiant. Probably Approximately [Yampolskiy, 2013] Roman V. Yampolskiy. Turing Test as
Correct: Nature's Algorithms for Learning and Prosper- a Defining Feature of AI-Completeness. Artificial
ing in a Complex World, Basic Books (AZ). Intelligence, Evolutionary Computation and
[Vinyals et al., 2019] Oriol Vinyals, Igor Babuschkin, Metaheuristics - In the footsteps of Alan Turing. Xin-She
Wojciech M Czarnecki, Michaël Mathieu, Andrew Yang (Ed.), Springer: 3-17.
Dudzik, Junyoung Chung, David H Choi, Richard Pow- [Yudkowsky and Hanson, 2008] Eliezer Yudkowsky and
ell, Timo Ewalds and Petko Georgiev. "Grandmaster Robin Hanson. The Hanson-Yudkowsky AI-foom
level in StarCraft II using multi-agent reinforcement debate. MIRI Technical Report. Available at:
learning." Nature: 1-5. http://intelligence.org/files/AIFoomDebate.pdf.
[Wang, 2019] Pei Wang. "On Defining Artificial Intelli- [Zhuang et al., 2009] Li Zhuang, Feng Zhou and J Doug
gence." Journal of Artificial General Intelligence 10(2): Tygar. "Keyboard acoustic emanations revisited." ACM
1-37. Transactions on Information and System Security
[Wiggers, May 2, 2020] Kyle Wiggers. Yann LeCun and (TISSEC) 13(1): 1-26.
Yoshua Bengio: Self-supervised learning is the key to [Ziesche and Yampolskiy, 2020] Soenke Ziesche and
human-level intelligence. Available at: https://venture- Roman V Yampolskiy. "Towards the Mathematics of
beat.com/2020/05/02/yann-lecun-and-yoshua-bengio- Intelligence." The Age of Artificial Intelligence: An
self-supervised-learning-is-the-key-to-human-level-in- Exploration: 1.
telligence/.
[Wigner, 1990] Eugene P Wigner. The unreasonable
effectiveness of mathematics in the natural sciences.
Mathematics and Science, World Scientific: 291-306.
[Wolpert, 2012] David H Wolpert. What the no free lunch
theorems really mean; how to improve search
algorithms. Santa Fe Institute.
[Wolpert and Macready, 1997] David H Wolpert and
William G Macready. "No free lunch theorems for
optimization." IEEE transactions on evolutionary
computation 1(1): 67-82.
[Yaman et al., 2020] Dogucan Yaman, Fevziye Irem
Eyiokur and Hazım Kemal Ekenel. "Ear2Face: Deep
Biometric Modality Mapping." arXiv preprint
arXiv:2006.01943.
[Yampolskiy, 2012] Roman V Yampolskiy. "AI-Complete,
AI-Hard, or AI-Easy–Classification of Problems in AI."
The 23rd Midwest Artificial Intelligence and Cognitive
Science Conference, Cincinnati, OH, USA.
[Yampolskiy, 2015] Roman V Yampolskiy. The space of
possible mind designs. International Conference on
Artificial General Intelligence, Springer.
[Yampolskiy, 2017] Roman V Yampolskiy. "What are the
ultimate limits to computational techniques: verifier
theory and unverifiability." Physica Scripta 92(9):
093001.