Extended Summary of "Living Things Are Not (20th Century) Machines: Updating Mechanism Metaphors in Light of the New Science of Machine Behavior"

EXTENDED SUMMARY OF “Living Things Are Not
(20th
Century) Machines: Updating Mechanism Metaphors
in Light of the Modern Science of Machine Behavior” by
Joshua Bongard and Michael Levin (2021)
LAUREANDO: RELATORE:
Simone Cappiello Eric Medvet
MATRICOLA: IN0500523
Anno Accademico 2021/2022
INTRODUCTION
M. Levin and J. Bongard’s main purpose with this paper is to highlight the need
to revise the old, widely used in scientific context and beyond, “life as
machine” metaphor. The authors examine 20th
century definitions that set
clear boundaries between machines and life, update them, for they came from
technological limitations now overcome, and identify more essential aspects
defining artificial and biological agents. They also argue that it will be
impossible to classify and better comprehend the emerging, increasingly
complex, hybrid systems (made of both artificial and biological components,
both evolved and designed) without questioning our understanding of these
basic concepts. Lastly, they emphasize the rise of a new multidisciplinary field,
mixing aspects of biology, information sciences and physics, based on
embodied computation. Thus, they look for an improved definition of
“machine” that also applies to biology and can facilitate scientific research.

MACHINE/LIFE DUALISM
The authors revise 20th
century statements about machines to prove that most
of the limiting features that used to clearly distinguish them from biological
systems are fading away.
-“Machines are independent, life is interdependent”: to prove this wrong,
examples of biohybrid machines for sensory/performance augmentation and
brain-computer interfaces are presented where communication and
cooperation of these devices with the environment in which they are placed
(typically a biological body) is crucial[1][2]. Then they point out that modern
machines are clearly made of complex interoperating components that are
often made of smaller parts that depend and rely on each other to properly
work.
-“Machines are predictable and designed, Life is unpredictable and evolved”:
No longer true because machines too are showing unexpected behavior
(perverse instantiation) especially in experiments with evolutionary algorithms,
specific kinds of bioinspired machine learning algorithms [3]. Evolutionary
algorithms have also proved that robots [4], synthetic organisms [5], jet
engines [6], and other machines can be evolved.
-“Life is hierarchical and self-similar, Machines are linearly modular”:
Living creatures are made of parts that exhibit specific competencies at
different scales and can adapt to change [7]. Self-similarity comes with
autonomy at different scales. Tadpoles, for instance, can cooperate in swarms
while competing to achieve their individual goals. The same pattern of
cooperation/competition is found in their tissues, which compete for nutrients
and information [8], and, at smaller scales, in their cells. This quality of
biological systems maximizes problem solving, robustness and adaptability
[9][10]. On the other hand, though machines, like living things, are modular
(made of interoperating and communicating subsystems, modules), they lack
self-similarity: their components can’t be conceived as individuals that seek
their own goals. Therefore, authors admit this is still a consistent statement,
even though they hypothesize that efforts to implement self-similarity will
accelerate due to its evolutionary advantages [8].

-“Life is capable of intelligence, Machines are not and never will”:
The authors argue that the origins of intelligence and its functions
(metacognition, consciousness, subjectivity, free will etc.) are still unclear [11],
and debates about what cognition is and does are still open [12]. One
profitable way to classify the intelligence of a system is its grade of
persuadability, the ability to change behavior through low-energy
interventions (messages, words) without the need of physical pushes (rewiring
or replacing parts, for example). The higher the grade of persuadability, the
lower the energy needed to influence the actions of the system. Modern
autonomous systems’ persuadability level is increasing, therefore machines’
cognitive capacities are expanding.
Biohybrid systems like living cells interfaced with optogenetic and machine
learning architectures that control their behavior [13][14] falsify the idea that
there is a way to sharply distinguish systems that exhibit subjectivity from the
ones that don’t.
-“Machines can be studied in a reductionist framework, Life cannot”:
With the rise of complexity of AI systems, reductionist analysis (understanding
the way a system functions by looking at its components) is losing its
efficiency. The authors state that it should not be surprising that technologies
like neural networks and swarm AI, respectively based on the biological
nervous system and swarms of animals, need the same approach used on their
biological counterparts to be understood [15][16]. In many cases, machines,
even the simplest ones [17], resist reductionist analysis and are better
understood through a top-down/high level (focused on their memories,
motivations, goals, beliefs) analysis. Hence, a new science studying intelligent
machines as active agents in relation to a specific context, machine behavior, is
emerging [18]. Machine behavior uses methods drawn from ethology,
cognitive and social sciences to predict and understand AI systems’ actions.
-“Life is embodied and doesn’t have clear hardware/software distinction, AIs
are not embodied and have clear hardware/software distinction”:
Though machine learning algorithms that run on robots seem to highlight a
hardware/software dualism, there are experiments where robots evolve along
with the software they run and learn to model their own body and what
happens to it (movement or damage, for example) [19]. These situations blur
the distinction between embodied robots and software AI algorithms by
merging the two in a single entity. The distinction between hardware and

software is blurry in both biology and technology. The authors cite examples of
objects that are commonly considered hardware influencing software and the
way systems work (we’d expect the opposite: hardware that is controlled by
software). It's reasonable to identify electrical activity in the brain as software
and cells as hardware in biology. The authors confute this hypothesis showing
that even blood flow and neurotransmitters (physical entities that would be
identified as hardware) can carry information. The same distinction is arbitrary
in technology. For example, inconvenient body dynamics of soft machines can
be used as computational resources [20]. Applications of DNA computing [21]
and robots built from DNA [22] fade this distinction furthermore.
IMPROVING DEFINITIONS
The authors here update definitions of machine, robot, program,
hardware/software specifying their intention to stimulate discussions, open
new unasked research questions and unify research programs that are
mistakenly considered distinct. Here are their definitions’ updates:
-Machine:
A system that uses principles of physics and computation to amplify the power
of an agent to make changes in its environment. Machines can be both evolved
and designed. Their behavior can be modified by interventions at physical level
or at higher level through communication/messages. They’re not necessarily
physical entities.
-Robot:
Machine that affects its environment through physical action. The degree of
roboticism should be a spectrum with autonomy from human control and
persuadability as two important parameters to classify robots in the spectrum:
the more autonomous they are and the smaller the energy required to change
their long-term behavior, the higher the degree of roboticism.

-Program:
Defined as an abstract procedure that can be executed and realized in many
ways. The same program can be executed on different physical systems that
support computation, with no restrictions on the medium that carries and the
medium that executes the information. Given that programs neither need to
be written by humans nor be linear one-step-at-a-time instructions, this notion
is extended to biological systems.
Software/Hardware:
Two interpretations based on etymology are presented:
1.Software as the flow of electrons/photons (whatever carries the information)
through circuitry opposed to hardware (the components of the circuits, like
metal, transistors etc).
2. Hardware as harder to modify/repair than software.
The latter assumption was overturned in one study about a soft robot [23], the
first one is poorly consistent for reasons we already explained. Thus, the
authors point out that taking for granted a sharp distinction between software
and hardware might not be useful in creating intelligent machines and studying
biological adaptation.
HYBRID SYSTEMS AND THE MULTIDISCIPLINARY BENEFITS
OF A NEW SCIENCE OF MACHINES
Hybridization:
Hybridization here is presented in more detail citing systems like insect-
machine hybrid robots [24] and robot gardens [25]: applications where
biological systems and electronic components strictly cooperate, making these
entities difficult to categorize. Because of the increasing hybridization of
systems, they propose a new way to categorize existing entities, looking at
categorization as a continuum in which a system can be partly biological and
partly inorganic as shown in Figure 1 (Source: Bongard and Levin 2021).

Figure 1 : Multi-axis categorization spectrum
Source: Bongard and Levin 2021
Systems are placed on a multi-scale axis based on the level of organization
(from cells, to individuals, to swarms, z axis), degree of autonomy (from merely
mechanical to high-level cognition) and degree of design (from designed to
evolved systems).

Benefits of Machine Behavior as a new science:
The authors list the advantages brought by the study of machine behavior to
related research topics. Some of them are:
-Our better knowledge of unpredictable and multi-scale systems will improve
techniques of reverse-engineering and multi scale analysis that are particularly
useful in regenerative medicine and developmental biology [10].
-Facing the problem of the implementation of agency, motivation, seeking and
setting goals in synthetic systems will clarify which features of living things are
the sources and causes of these capacities.
-Biomimicry, inspiration from functions of biological systems, is especially
useful in artificial intelligence research. Notable examples of applications of
this method are convolutional neural networks and deep reinforcement
learning [26][27].
CONCLUSION
According to the authors, biology and computer science must be seen as
strictly correlated study fields, both subsets of information sciences, dealing
with similar issues in different media. A better cooperation between these two
science fields will open new perspectives in empirical research and will
improve our conceptual understanding of agency, computation, cognition and
all the fundamental activities that biological, synthetic and hybrid agents
perform. This path will widen the space of possible embodied computing
systems (biological, artificial or both: placed in the multi scale spectrum of
hybrids previously shown) and help them in attaining their full potential and
utility.

REFERENCES
[1] Sampaio, E., Maris, S., and Bach-y-Rita, P. (2001). Brain plasticity: ‘visual’ acuity of blind
persons via the tongue. Brain Res. 908, 204–207. doi: 10.1016/S0006-8993(01)02667-1.
[2] Shanechi, M. M., Hu, R. C., and Williams, Z. M. (2014). A cortical-spinal prosthesis for
targeted limb movement in paralysed primate avatars. Nat. Commun. 5:3237. doi:
10.1038/ncomms4237.
[3] Lehman, J., Clune, J., Misevic, D., Adami, C., Altenberg, L., and Beaulieu, J. (2020). The
surprising creativity of digital evolution: a collection of anecdotes from the evolutionary
computation and artificial life research communities. Artif. Life 26, 274–306.
[4] Brodbeck, L., Hauser, S., and Iida, F. (2018). “Robotic invention: challenges and perspectives
for model-free design optimization of dynamic locomotion robots,” in Robotics Research, eds A.
Bicchi and W. Burgard (Cham: Springer), 581–596. doi: 10.1007/978-3-319-60916-4_33.
[5] Kriegman, S., Blackiston, D., Levin, M., and Bongard, J. (2020). A scalable pipeline for
designing reconfigurable organisms. Proc. Natl. Acad. Sci. U.S.A. 117, 1853–1859. doi:
10.1073/pnas.1910837117.
[6] Yu, X., Wang, C., and Yu, D. (2019). Configuration optimization of the tandem cooling-
compression system for a novel precooled hypersonic airbreathing engine. Energy Convers.
Manag. 197:111827. doi: 10.1016/j.enconman.2019.111827.
[7] Schulkin, J., and Sterling, P. (2019). Allostasis: a brain-centered, predictive mode of
physiological regulation. Trends Neurosci. 42, 740–752. doi: 10.1016/j.tins.2019.07.010.
[8] Gawne, R., McKenna, K. Z., and Levin, M. (2020). Competitive and coordinative interactions
between body parts produce adaptive developmental outcomes. BioEssays 42:e1900245. doi:
10.1002/bies.201900245.
[9] Levin, M. (2019). The computational boundary of a “self”: developmental bioelectricity drives
multicellularity and scale-free cognition. Front. Psychol. 10:2688. doi:
10.3389/fpsyg.2019.02688.
[10] Levin, M. (2020). Life, death, and self: fundamental questions of primitive cognition viewed
through the lens of body plasticity and synthetic organisms. Biochem. Biophys. Res.
Commun. doi: 10.1016/j.bbrc.2020.10.077 .
[11] Lyon, P. (2006). The biogenic approach to cognition. Cogn. Process. 7, 11–29. doi:
10.1007/s10339-005-0016-8.

[12] Bronfman, Z. Z., Ginsburg, S., and Jablonka, E. (2016). The transition to minimal
consciousness through the evolution of associative learning. Front. Psychol. 7:1954.
[13] Newman, J. P., Fong, M. F., Millard, D. C., Whitmire, C. J., Stanley, G. B., and Potter, S. M.
(2015). Optogenetic feedback control of neural activity. Elife 4:e07192.
[14] Grosenick, L., Marshel, J. H., and Deisseroth, K. (2015). Closed-loop and activity-guided
optogenetic control. Neuron 86, 106–139. doi: 10.1016/j.neuron.2015.03.034.
[15] Valentini, G., Moore, D. G., Hanson, J. R., Pavlic, T. P., Pratt, S. C., et al. (2018). “Transfer
of information in collective decisions by artificial agents,” in Proceedings of the 2018 Conference
on Artificial Life, (Cambridge, MA: MIT press), 641–648.
[16] Beer, R. D., and Williams, P. L. (2015). Information processing and dynamics in minimally
cognitive agents. Cogn. Sci. 39, 1–38. doi: 10.1111/cogs.12142.
[17] Jonas, E., and Kording, K. (2016). Could a neuroscientist understand a
microprocessor? biooRxiv [preprint] doi: 10.1371/journal.pcbi.1005268.
[18] Rahwan, I., Cebrian, M., Obradovich, N., Bongard, J., Bonnefon, J. F., and Breazeal, C.
(2019). Machine behaviour. Nature 568, 477–486.
[19] Kwiatkowski, R., and Lipson, H. (2019). Task-agnostic self-modeling machines. Sci.
Robot. 4:eaau9354. doi: 10.1126/scirobotics.aau9354.
[20] Nakajima, K., Hauser, H., Li, T., and Pfeifer, R. (2015). Information processing via physical
soft body. Sci. Rep. 5:10487.
[21] Chatterjee, G., Dalchau, N., Muscat, R. A., Phillips, A., and Seelig, G. (2017). A spatially
localized architecture for fast and modular DNA computing. Nat. Nanotechnol. 12, 920–927. doi:
10.1038/nnano.2017.127.
[22] Thubagere, A. J., Li, W., Johnson, R. F., Chen, Z., Doroudi, S., and Lee, Y. L. (2017). A
cargo-sorting DNA robot. Science 357:eaan6558.
[23] Shah, D. S., Powers, J. P., Tilton, L. G., Kriegman, S., Bongard, J., and Kramer-Bottiglio, R.
(2020). A soft robot that adapts to environments through shape change. Nat. Mach. Intell. 3, 51–
59. doi: 10.1038/s42106-020-00263-1.
[24] Ando, N., and Kanzaki, R. (2020). Insect-machine hybrid robot. Curr. Opin Insect. Sci. 42,
61–69. doi: 10.1016/j.cois.2020.09.006.
[25] von Mammen, S., Hamann, H., and Heider, M. (2016). “Robot gardens: an augmented
reality prototype for plant-robot biohybrid systems,” in Proceedings of the 22nd ACM Conference
on Virtual Reality Software and Technology, ed. E. Kruijff (New York, NY: Association for
Computing Machinery), 139–142. doi: 10.1145/2993369.2993400.
[26] Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2017). ImageNet classification with deep
convolutional neural networks. Commun. ACM 60, 84–90. doi: 10.1145/3065386.
[27] Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., and Bellemare, M. G. (2015).
Human-level control through deep reinforcement learning. Nature 518, 529–533.

Extended Summary of "Living Things Are Not (20th Century) Machines: Updating Mechanism Metaphors in Light of the New Science of Machine Behavior"

More Related Content

Similar to Extended Summary of "Living Things Are Not (20th Century) Machines: Updating Mechanism Metaphors in Light of the New Science of Machine Behavior" (20)

Recently uploaded (20)

Extended Summary of "Living Things Are Not (20th Century) Machines: Updating Mechanism Metaphors in Light of the New Science of Machine Behavior"