The Cortical Conductor Theory: Towards

Addressing Consciousness in AI Models

Joscha Bach(&)

Harvard Program for Evolutionary Dynamics, Cambridge, MA 02138, USA

[email protected]

Abstract. AI models of the mind rarely discuss the so called “hard problem” of
consciousness. Here, I will sketch informally a possible functional explanation
for phenomenal consciousness: the conductor theory of consciousness (CTC).
Unlike IIT, CTC is a functionalist model of consciousness, with similarity to
other functionalist approaches, such as the ones suggested by Dennett and
Graziano.

Keywords: Phenomenal consciousness
Cortical conductor theory

Attention
Executive function
Binding
IIT

1 Introduction: Consciousness in Cognitive Science

While AI offers a large body of work on agency, autonomy, motivation and affect,
cognitive architectures and cognitive modeling, there is little agreement on how to
address what is usually called “the hard problem” of consciousness. How is it possible
that a system can take a first person perspective, and have phenomenal experience?
One of the better known recent attempts to address phenomenal consciousness is
Guilio Tononi’s Integrated Information Theory (IIT) [1, 2]. Tononi argues that expe-
rience cannot be reduced to a functional mechanism, and hence it must be an intrinsic
property of a system, rather than a functional one. He characterizes consciousness by a
parameter, U, which is a measure for the amount of mutual information over all
possible partitionings of an information processing system. If the information in the
system is highly integrated (i.e. the information in each part of the system is strongly
correlated with the information in the others), it indicates a high degree of con-
sciousness. As has for instance been argued by Aaronson [3], IIT’s criterion of
information integration could perhaps be necessary, but is not sufficient, because we
can construct structurally trivial information processing systems that maximize U by
maximally distributing information (for instance via highly interconnected XOR gates).
Should we assign consciousness to processing circuits that are incapable of exhibiting
any of the interesting behaviors of systems that we usually suspect to be conscious,
such as humans and other higher animals?
From a computationalist perspective, IIT is problematic, because it suggests that
two systems that compute the same function by undergoing a functionally identical
sequence of states might have different degrees of consciousness based on the
arrangement of the computational elements that realize the causal structure of the

© Springer Nature Switzerland AG 2019

A. V. Samsonovich (Ed.): BICA 2018, AISC 848, pp. 16–26, 2019.
https://doi.org/10.1007/978-3-319-99316-4_3
 The Cortical Conductor Theory 17

system. A computational system might turn out to be conscious or unconscious

regardless of its behavior (including all its utterances professing its phenomenal
experience) depending on the physical layout of its substrate, or the introduction of a
distributed virtual machine layer. A more practical criticism stems from observing
conscious and unconscious people: a somnambulist (who is generally not regarded as
conscious) can often answer questions, navigate a house, open doors etc., and hence
should have cortical activity that is distributed in a similar way as it is in an awake,
conscious person [4]. In this sense, there is probably only a low quantitative difference
in U, but a large qualitative difference in consciousness. This qualitative difference can
probably be explained by the absence of very particular, local functionality in the brain
of the somnambulist: while her cortex still produces the usual content, i.e. processes
sensory data and generates dynamic experiences of sounds, patterns, objects, spaces
etc. from them, the part that normally attends to that experience and integrates it into a
protocol is offline. This integrated experience is not the same as information integration
in IIT. Rather, it is better understood as a particular local protocol by one of the many
members of the “cortical orchestra”: its conductor.
In this contribution, I will sketch how a computational model can account for the
phenomenology and functionality of consciousness, based on my earlier work in the
area of cognitive architectures [5]; we might call this approach the “conductor theory of
consciousness” (CTC).

2 An AI Perspective on the Mind

Organisms evolved information processing capabilities to support the regulation of

their systemic demands in the face of the disturbances by the environment. The sim-
plest regulator system is the feedback loop: a system that measures some current value
and exerts a control operation that brings it close to a target value. Using a second
feedback loop to regulate the first, the system can store a state and regulate one value
depending on another. By changing the control variables to maximize a measured
reward variable, a system can learn to approximate a complex control function that
maps the values of a set of inputs (sensors) to operators (effectors). Our nervous
systems possess a multitude of feedback loops (such as the mechanisms of the brain
stem regulating heart rate and breathing patterns).
The control of behavior requires more complex signals; the sensors of the limbic
system measure changes in organismic demands and respond to satisfaction of needs
with pleasure signals (indicating to intensify the current behavior). The frustration of
needs leads to displeasure signals (pain), which indicate the current behavior should be
stopped. Directed behavior of a system may be governed by impulses, which associate
situations (complex patterns in the outer or inner environment of the organism) with
behavior to obtain future pleasure, or avoid future pain. Pain and pleasure act as reward
signals that establish an association between situations, actions and needs [6]. In
mammals, such connections are for instance established in the hippocampus [7]
(Fig. 1).
18 J. Bach

Fig. 1. From needs to cognition

The human neocortex enables better regulation of needs by encoding sensory

patterns into a complex hierarchical model of the environment (including the inner
environment). This dynamic model is not just a mapping from past observation to
future observations, but takes on the shape of a progressively updated stateful function,
a program that generates a simulation of the environment. The formation of the model
is driven largely by data compression, i.e. by optimizing for the data structure that
allows the best predictions of future observations, based on past observations. This
principle has for instance been described by Solomonoff [8]: The best possible model
that a computational agent can form about its environment is the shortest program
among those that best predict an observation from past observations, for all observa-
tions and past observations. Machine learning models of the mind can be understood as
approximating Solomonoff induction [9], by capturing the apparent invariances of the
world into an almost static model, and its variance as a variable state of that model. By
varying the state, such a model cannot only capture the current state of the world, but
be used to anticipate and explore possible worlds, to imagine, create and remember.
Machine learning systems have demonstrated how recurrent neural networks can dis-
cover and predict the structure of visual and auditory stimuli by forming low level
feature detectors, which can then be successively organized into complex high level
 The Cortical Conductor Theory 19

features, object categories and conceptual manifolds [10]. Deep networks can form
hierarchical knowledge representations. LSTMs [11] and GRUs [12] are building
blocks for recurrent neural networks that can learn sequences of operations. Generative
neural networks can use the constraints learned from the data to produce possible
worlds [13].
While current machine learning systems outperform humans in many complex
tasks that require the discovery and manipulation of causal structures in large problem
spaces, they are very far from being good models of intelligence. Part of this is due to
our current learning paradigms, which lead to limitations in the generation of com-
positional knowledge and sequential control structures, and will be overcome with
incremental progress.
Recently, various researchers have proposed to introduce a unit of organization
similar to cortical columns into neural learning [14]. Cortical columns are elementary
circuits containing between 100 and 400 neurons [15], and are possibly trained as echo
state networks [16] to achieve functionality for function approximation, conditional
binding and reward distribution. In the human neocortex, the columnar units form
highly interconnected structures with their immediate neighbors, and are selectively
linked to receptive fields in adjacent cortical areas. A cortical area contains ca. 106 to
107 columns, and may be thought of as a specialized instrument in the orchestra of the
neocortex.
Beyond Current Machine Learning
A more important limitation of many current machine learning paradigms is their
exclusive focus on policy learning and classification. Our minds are not classifiers—
they are simulators and experiencers. Like machine learning systems, they successively
learn to identify features in the patterns of the sensory input, which they then combine
into complex features, and organize into maps. High-level features may be integrated
into dynamic geometries and objects, motor patterns and procedures, auditory structure
and so on. Features, objects and procedures are sensory-motor scripts that allow the
manipulation of mental content and the execution of motor actions.
Unlike most machine learning systems, our minds combine these objects, maps and
procedural dynamics into a persistent dynamic simulation, which can be used to
continuously predict perceptual patterns at our systemic interface to the environment
(Fig. 2). The processing streams formed by the receptive fields of our cortical instru-
ments enable the bottom-up cuing of perceptual hypotheses (objects, situations etc.),
and trigger the top-down verification of these hypotheses, and the binding of the
features into a cohesive model state. The elements of this simulation do not necessarily
correspond to actual objects in the universe: they are statistical regularities that our
mind discovered in the patterns at its systemic interface. Our experience is not directed
on the pattern generator that is the universe, but on the simulation produced in our
neocortex. Thus, our minds cannot experience and operate in an “outer” reality, but in a
dream that is constrained to past and current sensory input [17].
20 J. Bach

Fig. 2. Gradual abstraction from sensory patterns to mental simulations

Human cognition does not stop at generative simulations, however. We can abstract
our mental representations into a conceptual manifold (Fig. 3). Concepts can be
thought of as an address space for our sensory-motor scripts, and they allow the
interpolation between objects, as well as the manipulation and generation of previously
unknown objects via inference. The conceptual manifold can be organized and
manipulated using grammatical language, which allows the synchronization of con-
cepts between speakers, even in the absence of corresponding sensory-motor scripts.
(The fact that language is sufficient to infer the shape of the conceptual manifold
explains the success of machine translation based on the statistical properties of large
text corpora, despite the inability of these systems to produce corresponding mental
simulations.)

Fig. 3. Conceptual abstraction

 The Cortical Conductor Theory 21

3 The Cortical Conductor

Cortical columns may be thought of as elementary agents that self-organize into the
larger organizational units of the brain areas as a result of developmental reinforcement
learning. The activity of the cortical orchestra is highly distributed and parallelized, and
cannot be experienced as a whole. However, its performance is coordinated by a set of
brain areas that act as a conductor. The conductor is not a “homunculus”, but like the
other instruments, a set of dynamic function approximators. Whereas most cortical
instruments regulate the dynamics and interaction of the organism with the environ-
ment (or anticipated, reflected and hypothetical environments), the conductor regulates
the dynamics of the orchestra itself. Based on signals of the motivational system, it
provides executive function (i.e. determines what goals the system commits to at any
given moment), resolves conflicts between cortical agents, and regulates their activa-
tion level and parameterization. Without the presence of the conductor, our brain can
still perform most of its functions, but we are sleep walkers, capable of coordinated
perceptual and motor action, but without central coherence and reflection (Fig. 4).

Fig. 4. The cortical conductor

In the human brain, the functionality of the conductor is likely facilitated via the
dorsolateral prefrontal cortex [18–20], anterior cingulate cortex and anterior insula
[21]. The conductor has attentional links into most regions. In each moment, it directs
its attention to one or a few of the cortical instruments, while others continue to play
unobserved in the background. The conductor may not access the activity of the region
it attends to in its entirety, but it may usually access some of the currently relevant
processing states and parameters of it. To learn and to reflect, the conductor maintains a
protocol of what it attended to, as a series of links to experiences generated by the other
22 J. Bach

cortical instruments. This protocol may be used to address the currently active regions,
and to partially recreate past states of the mental simulation by reactivating the cor-
responding configuration of active regions with the parameters of the stored links. The
reactivation of a past state of the mental simulation will generate a re-enactment of a
previous world state: a memory. Further abstraction of the protocol memory leads to
the formation of new kinds of sensory motor scripts: an autobiographical memory
(events that happened to the agent), and a procedural memory. The reflective access to
the protocol allows learning and extrapolation of past events, and the act of accessing
the protocol may of course itself become part of the protocol. By accessing the memory
of the access to its own protocol, the system remembers having had access to expe-
rience (access consciousness).
While all cortical regions store information as a result of updating their models and
learning associations to motivational signals, the attentional protocol of the conductor
is the only place where experience is integrated. Information that is not integrated in the
protocol cannot become functionally relevant to the reflection of the system, to the
production of its utterances, the generation of a cohesive self model, and it cannot
become the object of access consciousness.
Phenomenal consciousness may simply be understood as the most recent memory
of what our prefrontal cortex attended to. Thus, conscious experience is not an expe-
rience of being in the world, or in an inner space, but a memory. It is the reconstruction
of a dream generated more than fifty brain areas, reflected in the protocol of a single
region. By directing attention on its own protocol, the conductor can store and recreate
a memory of its own experience of being conscious.
The idea that we are not actually conscious in the moment, but merely remember
having been conscious is congruent with known inconsistencies in our experience of
consciousness, such as subjective time dilation, false continuity, and loops in the
conscious experience. Subjective dilation of time results from states of high arousal, for
instance during an accident, whereas uneventful flow states often lead to a subjective
contraction of time. Both dilated and contracted time do not correspond to an increase
or decrease in the actual processing speed of our cognitive operations. Instead, they
result from a higher or lower number of entries in the protocol memory: the experi-
enced time interval only seems to be longer or shorter with hindsight. An extreme case
of a subjective dilation of time can happen during dreams, which sometimes play out in
a physical time interval of a few seconds of REM sleep, yet may span hours of
subjective time. This may be explained by the spontaneous generation of the entire
dream, rather than the successive experience of each event. Hour-long dreams are
probably simply false memories.
False continuity results from gaps in our attention, for instance during saccadic
movements, or interruptions and distractions of gaze. While these breaks in attention
may lead to missing significant changes in parts of the environment that we believe we
are attending to, they are not part of the protocol memory and hence our experience
appears to be unbroken in hindsight. For a considerable fraction of our days, we are
probably wakeful but not conscious. Inconsistent experiences of consciousness can be
explained as false memories, but they do not have subjective qualities that makes them
appear “less conscious” than consistent experiences. Thus, if at least some of our
conscious experience is a false memory, why not all of it? Treating consciousness as a
 The Cortical Conductor Theory 23

memory instead of an actual sense of the present resolves much of the difficulty for
specifying an AI implementation of consciousness: it is necessary and sufficient to
realize a system that remembers having experienced something, and being able to
report on that memory.

4 Consciousness and Self Model

In the above discussion, I have treated phenomenal consciousness in the sense of “the
feeling of what it’s like”. However, consciousness is often associated with more
concrete functionality, especially a specific model of self, and a set of functionality
pertinent to that model. This has lead Marvin Minsky to call consciousness “a suitcase
term” [22], a notion that is notoriously hard to unpack.
Conscious states differ by the configuration and available functionality of a cog-
nitive system at a given time. However, once we understand how an attentional pro-
tocol can provide for binding of other cortical functionality into a single structure for
the purpose of self regulation, we can enumerate some of the functionality that cor-
responds to a given conscious state.
Core consciousness is characterized by:
– a local perceptual space
– the ability to access mentally represented percepts
– a current world model
– directed attention (inwards/outwards, wide/focused)
– the ability to access and follow concepts and similar content
– the ability to manipulate and create concepts and similar content,
– the presence of an inner stage of currently active, non-perceptual concepts and
associative representations
In deep meditation, the following functionality may be absent:
– an integrated personal self-model (sense of identity)
– a sense of one’s own location and perspective in space
– proprioception (position and state of body and limbs)
– valences (pleasure and displeasure signals)
– goals and committed plans
– the awareness of the current affective state
– the influence of desires and urges on behavior
– the ability to create and process discourse (i.e. translate mental representations into
communicable symbols, and vice versa)
Lucid dreams are specific dream states that are different from wakefulness by the
absence of:
– access to needs/desires, urges
– access to sensory perception
– the ability to exert voluntary control over muscles
– a biographical memory and protocol
24 J. Bach

– a short term biography

– the ability to separate perceptually grounded content from ideas/imaginations
Dreams are usually in addition characterized by the absence of:
– having accessible knowledge about the fact that there access to percepts and con-
cepts (access consciousness)
– a social model of self (self-ascription of beliefs, desires, intentions, skills, traits,
abilities, personality)
– the formation of and access to expectations of immediate future
– the ability to influence behavior based on discursive thought
– the ability to relate self-ascribed actions to apparent mental causes (sense of agency)
– the ability to form memories of the current content
– the ability to reason
– the ability to construct plans
– the ability to act on plans
– the ability to keep goals stable until they are achieved
– the ability to let go of goals that are unattainable
– the ability to signal aspects of one’s mental state to others
Diminished states of consciousness (for instance, in small children or due to neu-
rodegenerative diseases) may also impair:
– the ability to influence behavior based on past experience (learning)
– the ability to construct causal models of the environment
– the ability to construct intentional models of agents
The above functionality is part of the general functionality of a human-like cognitive
agent and has to be implemented into its cognitive architecture, either explicitly or via a
self-organized process of reward driven learning (each of them can be realized on
computational machines). The differences between conscious states result from the
dissociation or impairment of these functions.
Is the conductor a learned or a predefined structure? I suspect that the formation of
the conductor functionality is itself a process of developmental learning, driven by
rewards for the self-regulation of cognition, and developmental cues that regulate the
onset and some of the parameters of the formation of the structure. Multiple personality
disorder lends further credibility to the hypothesis that the conductor is constructed by
reward driven neural self-organization. In patients with multiple personalities, the
different personas usually do not share a subjective protocol, biographical and proce-
dural memory. But even if we form multiple conductors, they share infrastructure (such
as linguistic processing, access to the attentional network and information transfer via
the thalamic loop), which ensures that only one of them may be online and form
memories at any given moment.
 The Cortical Conductor Theory 25

5 Summary

The cortical conductor theory (CTC) posits that cortical structures are the result of
reward driven learning, based on signals of the motivational system, and the structure
of the data that is being learned. The conductor is a computational structure that is
trained to regulate the activity of other cortical functionality. It directs attention, pro-
vides executive function by changing the activity and parameterization and rewards of
other cortical structures, and integrates aspects of the processes that it attended to into a
protocol. This protocol is used for reflection and learning. Memories can be generated
by reactivating a cortical configuration via the links and parameters stored at the
corresponding point in the protocol. Reflective access to the protocol is a process that
can itself be stored in the protocol, and by accessing this, a system may remember
having had experiential access.
For phenomenal consciousness, it is necessary and sufficient that a system can
access the memory of having had an experience—the actuality of experience itself is
irrelevant (and logically not even possible).
CTC explains different conscious states by different functionality bound into the
self construct provided by the attentional protocol. The notion of integration is central
to CTC, however, integration is used in a very different sense than in Tononi’s Inte-
grated Information Theory (IIT). In CTC, integration refers to the availability of
information for the same cognitive process, within a causally local structure of an
agent. In IIT, integration refers to the degree in which information is distributed within
a substrate. CTC is a functionalist theory, and can be thought of as an extension to
Dennett’s “multiple drafts” model of consciousness [23]. CTC acknowledges that the
actual functionality of perception and cognition is distributed, disjoint and fragmentary,
but emphasizes the need to integrate access to this functionality for a module that in
turn has access to capabilities for reflection and the formation of utterances (otherwise,
there would be no self model and no report of phenomenal experience).
CTC also bears similarity to Michael Graziano’s attention schema theory of con-
sciousness [24]. Graziano suggests that just like the body schema models the body of
an agent, its attention schema models the activity and shape of its attentional network.
While the functionality subsumed under access consciousness, phenomenal con-
sciousness and conscious states, and the required mechanisms are slightly different in
CTC, we agree with the role of consciousness for shaping and controlling attention-
related mechanisms.

Acknowledgements. This work has been supported by the Harvard Program for Evolutionary
Dynamics, the MIT Media Lab and the Epstein Foundation. I am indebted to Katherine Gal-
lagher, Adam Marblestone, and the students of the Future of Artificial Intelligence course at the
MIT Media Lab for their contributions in discussions of the topic, as well as to Martin Novak and
Joi Ito for their support.
26 J. Bach

References
1. Tononi, G.: The integrated information theory of consciousness: an updated account. Arch.
Ital. Biol. 150, 56–90 (2012)
2. Tononi, G., Boly, M., Massimini, M., Koch, C.: Integrated information theory: from
consciousness to its physical substrate. Nat. Rev. Neurosci. 17(7), 450–461 (2016)
3. Aaronson, S.: Why I am not an integrated information theorist (or, the unconscious
expander) (2015). http://www.scottaaronson.com/blog/?p=1799. Accessed 15 Feb 2017
4. Zadra, A., Desautels, A., Petit, D., Montplaisir, J.: Somnambulism: clinical aspects and
pathophysiological hypotheses. Lancet Neurol. 12(3), 285–294 (2013)
5. Bach, J.: Principles of Synthetic Intelligence. Psi, An Architecture of Motivated Cognition.
Oxford University Press, Oxford (2009)
6. Bach, J.: Modeling motivation in MicroPsi 2. In: 8th International Conference on Artificial
General Intelligence, AGI 2015, Berlin, Germany, pp. 3–13 (2015)
7. Cer, D.M., O’Reilly, R.C.: Neural mechanisms of binding in the hippocampus and
neocortex. In: Binding in Memory. Oxford University Press, Oxford (2006)
8. Solomonoff, R.J.: A formal theory of inductive inference. Inf. Control 7, 224–254 (1964)
9. Hutter, M.: Universal Artificial Intelligence: Sequential Decisions Based on Algorithmic
Probability. EATCS Book. Springer, Cham (2005)
10. LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. Nature 521, 436–444 (2015)
11. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8), 1735–1780
(1997)
12. Cho, K., van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H.,
Bengio, Y.: Learning phrase representations using RNN encoder–decoder for statistical
machine translation. In: Proceedings of EMNLP, October 2014, pp. 1724–1734 (2014)
13. Dosovitskiy, A., Springenberg, J.T., Brox, T.: Learning to generate chairs with convolutional
neural networks. In: IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), pp. 1538–1546 (2015)
14. Hinton, G.E., Krizhevsky, A., Wang, S.D.: Transforming auto-encoders. In: International
Conference on Artificial Neural Networks. Springer, Berlin (2011)
15. Mountcastle, V.B.: The columnar organization of the neocortex. Brain 20(4), 701–722
(1997)
16. Jaeger, H.: Echo state network. Scholarpedia 2(9), 23–30 (2007)
17. Bach, J.: No room for the mind. Enactivism and artificial intelligence. In: Proceedings of the
Conference of History and Philosophy of Computing, Ghent (2011)
18. Bodovitz, S.: The neural correlate of consciousness. J. Theor. Biol. 254(3), 594–598 (2008)
19. Safavi, S., Kapoor, V., Logothetis, N.K., Panagiotaropoulos, T.I.: Is the frontal lobe
involved in conscious perception? Front. Psychol. 5, 1063 (2014)
20. Del Cul, A., Dehaene, S., Reyes, P., Bravo, E., Slachevsky, A.: Causal role of prefrontal
cortex in the threshold for access to consciousness. Brain 132(9), 2531–2540 (2009)
21. Fischer, D.B., Boes, A.D., Demertzi, A., Evrard, H.C., Laureys, S., Edlow, B.L., Liu, H.,
Saper, C.B., Pascual-Leone, A., Fox, M.D., Geerling, J.C.: A human brain network derived
from coma-causing brainstem lesions. Neurology 87(23), 2427–2434 (2016)
22. Minsky, M.: The Emotion Machine. Simon and Shuster, New York (2006)
23. Dennett, D.C.: Consciousness Explained. Back Bay Books, New York (1992)
24. Graziano, M.S.A., Webb, T.W.: A mechanistic theory of consciousness. Int. J. Mach.
Conscious. 6, 163–176 (2014)