The theory of functionalism holds that mental states are constituted solely by the functional organisation of an implementing system. We propose that the specific mechanism supporting both intelligence and consciousness is data compression. Recent approaches to cognition and artificial intelligence, based on a branch of theoretical computer science known as algorithmic information theory (AIT), have proposed a computational view of induction, prediction, understanding and consciousness which is founded on this concept. Building on existing literature we propose the term 'compressionism' as a label to unite theories which recognise intelligent, cognisant systems as sophisticated data compressors. We show how data compression can shed light on information binding and offer a novel perspective on the hard problem of consciousness.
Throughout history people have speculated whether the
human mind might be viewed as a sophisticated automaton.
The theory of functionalism holds that mental states are
constituted solely by their functional role
While functionalism tells us that intelligence and consciousness may be implemented in a number of ways, it does not bring us any closer to understanding the specific form of information processing that gives rise to these phenomena. Here, we explore the idea that the key function of the brain which supports both intelligence and consciousness is data compression. In the following sections we elucidate the nature of data compression and bring together existing theories which has highlighted its connection with intelligent thought and subjective experience. At first blush the concept ‘data compression’ might seem esoteric, a niche idea related to information storage in computer science. However, the concept runs much deeper. Data compression occurs when information is bound together through the identification of shared patterns. For example the sequence 4, 6, 8, 12, 14, 18, 20, 24... can be simplified as the description “odd prime numbers +1”. The latter representation is shorter, hence we can say that it has been ‘compressed’. The elegance and concision of this representation suggests that it is the true underlying pattern which governs the sequence. Accordingly, someone who manages to identify this pattern might claim to have understood the sequence, because they appreciate its relationship with the causal mechanism that produced it.
The above example highlights the close link between data
compression and prediction. As per Occam’s razor, concise
models make fewer assumptions and are thus more likely to
be correct.
Because of its connection to prediction, data compression
can be viewed as providing reliable proof of understanding.
According to
Take a simple drawing as another example. Without background knowledge and prior embodied experience, a computer program cannot associate the drawing with previously experienced objects. To encode the image it must individually register every pixel on the screen, resulting in a large data file, even when compression algorithms are applied. However, a human can recognise an even more complex pattern: for example, a drawing of a house. The pixels are not coloured randomly. Instead there is a clear pattern that explains their ordering and relations, e.g. the presence of a roof, windows and a door. These regularities can be summarised concisely without having to spell them out pixel by pixel.
The greater the extent of the identified patterns, the shorter the resulting representation and the better predictions it supports. Imagine that pixel (254, 56) is corrupted by noise. The uncompressed representation that stores each pixel individually cannot correct this error. However, if we can identify patterns in the data, we understand how it is connected. By appreciating this redundancy we can repair and filter errors. For instance, we realise that pixel (254, 56) is part of the roof of the house and thus should be coloured in the same way.
A weakness of the Turing test (Turing, 1950) is that a
program might pass the test simply by exploiting weaknesses in
human psychology. If a given system passes the test we
cannot be sure if it was because of the quality of the responses
or the gullibility of the judge. In contrast, Hutter’s
compression test is more reliable. The more that data is compressed,
the harder it becomes to compress it further
Research in the area of artificial intelligence and cognitive
science is increasingly identifying data compression as a key
organisational principle.
In a similar vein, both Maguire et al.’s (2013) theory of
subjective information and
Adopting the perspective of the mind as a compressor,
In the field of AI,
The connection between compression and prediction can provide an elegant explanation for why the brain has evolved to carry out compression so effectively. In order to thrive in an uncertain environment, organisms must be able to anticipate future events. The more efficiently they can compress their experiences, the more accurate their predictions will be. As a result, organisms have evolved brains which are prodigious compressors of information: compressing sensory information provides them with an understanding of their environment. We propose the term ‘compressionism’ to group together those theories which view data compression as the key organisational principle underlying the structure of intelligent thought. In the following sections we make the case that, as well as providing a grounding for a formal definition of intelligence, compressionism also provides the foundations for a philosophy of mind which can shed light on the combination problem and offer insight into the nature of subjective experience.
Intuitively, our subjective experience seems to carry the
qualitative characteristic of being united and singular, a property
which is at odds with the reducibility of the physical world.
How can the brain, whose processing is clearly distributed in
space and time, give rise to a consistently integrated
perspective?
Here James is highlighting the apparent incoherence of a collection of physical objects sharing a combined experience. We propose that combination is achieved, not by any local physical convergence or supernatural process, but through the process of data compression. When the brain compresses information, it is binding data together through the identification of shared patterns that might originally have been dispersed in space and time. The process of grouping and encoding these patterns yields computational results which reflect a ‘coming together’ of information. We will now explore the idea that this combination effect might be what underlies the united characteristic of conscious behaviour.
The reproductive success of an organism is dependent on cooperation between all of its constituent components, leading to a form of compression which unites data distributed across space and time. For instance, it does not make sense for an organism’s legs to maintain independent agenda. Because the interests of both legs are intimately bound, it is more productive for them to cooperate with each other in achieving a single set of objectives (e.g. walking). Accordingly, the brain sources sensory information from all over the body and binds it through compression, thereby optimising predictive accuracy for the organism as a whole. Tactile information from every limb is compressed alongside visual information from the eyes and audio information from the ears, giving rise to a form of understanding that is centralised and representative of the organism’s experiences as a singular unit. The resulting decisions of the organism also appear centralised: to the external observer it seems as if the organism’s body is being ‘controlled’ by a single entity.
The success of an organism also depends on cooperation
through time. Accordingly, the response it exhibits to a
sensory stimulus depends not just on its immediate processing,
but also on its memories. Patterns in a current stimulus
are matched against patterns distilled from historical
stimuli, leading to a form of understanding that unites not only
distributed sensory organs but also an organism’s past and
present states
The more memories we store, the greater our ability to identify and compress patterns in novel stimuli. For instance, when we observe an apple we can connect the sensory data with a large number of previously encountered apples, allowing us describe the stimulus concisely in terms of its relationship to an existing set of memories. Rather than needing to encode every detail of the fruit, we can simply note the ways in which it differs from a historical prototype, allowing greater redundancy to be identified, and thus enhancing predictive accuracy. In this way, data compression can support the identification of concepts.
Intuitively, features of the brain which are not linked to optimising behaviour should not have been rigorously preserved by evolution. The question thus arises of why the brain should go to the effort of producing consciousness at all. One advantage of the ‘compressionist’ perspective is that, rather than requiring an additional unique property of the brain to answer this question, it simply extends an approach which is already used to model intelligence. Given that it manages to link unitary processing to predictive and reproductive success, compressionism can provide an elegant explanation for why and how the brain has evolved to bind information.
The above account of binding through data compression bears close resemblance to Tononi’s (2008) integrated information theory. In line with compressionism, Tononi proposes that consciousness is an information processing phenomenon and can thus be quantified in terms of a system’s organisational structure, specifically its capacity to integrate information.
Tononi provides two thought experiments highlighting the need for conscious observations to a) produce information, and b) integrate information. In the first thought experiment he considers the difference between a human and a photodiode viewing a screen. The photodiode can only respond with two outputs, either light or dark. Accordingly, it generates a single bit of information about the stimulus, whereas a person, being able to distinguish between millions of different images, generates a much greater quantity of information. According to Tononi, the ability to discriminate between many different alternatives is an essential ingredient to conscious experience.
In his second thought experiment,
Intuitively, people’s conscious experiences cannot easily be broken down in this way. It would be far more difficult for a neurosurgeon to operate on somebody’s brain and alter their memory of a stimulus. Rather than being encoded in a discrete, localised region of the brain, as per the camera, a person’s sensory experience is integrated with a vast number of memories, causing widespread changes across all areas of the brain. Inside a digital camera the encoded information lies isolated, detached and dormant. In somebody’s brain the same information alters every aspect of how they think, act and process subsequent information.
Tononi’s (2008) idea of binding through information
integration is closely aligned with the idea of binding through
data compression. Indeed,
If data is optimally compressed then it becomes extremely
difficult to edit in its compressed state. For example, imagine
a compressed encoding of a Wikipedia page. You want to edit
the first word on the page. But where is this word encoded in
the compressed file? There is no easily delineated set of bits
which corresponds to the first word and nothing else. Instead,
the whole set of data has been integrated, with every bit from
the original file depending on all the others. To discern the
impact that the first word has had on the compressed
encoding you have to understand the compression scheme. There
are no shortcuts. Accordingly,
We propose that this unavoidable difficulty of breaking down people’s behaviour into a set of discrete, independent components is precisely what people mean when they apply the term ‘conscious’. From this integration or ‘compressionist’ perspective consciousness is not a tangible property, but rather a heuristic that people adopt in modelling the behaviour of a system. Specifically, we attribute consciousness to a system when the data compression it carries out is so sophisticated that we are forced to model it as unitary. In the case of the brain, its extraordinary capacity for compressing information leads us to attribute consciousness to the behaviour produced.
While compressionist theories of consciousness emphasise
integration as the defining feature of consciousness
Intuitively, consciousness does not appear ‘drafty’ (i.e.
easily decomposable into disjoint, unintegrated components).
While physical objects are clearly drafty, being reducible into
discrete atoms, the standout feature of conscious systems is
how undrafty they are. According to the compressionist
perspective, it is the undraftiness of a system’s information
processing we have in mind when we describe it as being
‘conscious’. Although
In the following sections we demonstrate the potential of compressionism as a theory of mind by detailing how it can account for other mysterious features of consciousness, such as qualia and self-awareness.
It seems possible to conceive of an artificial compressor which compresses large amounts of current and historical data in parallel, though without experiencing the same form of awareness that humans are familiar with. For example, critics of compressionism might question whether off-theshelf image or video compressors understand the images they are compressing. We argue that compression carried out by the brain is likely to have two additional ingredients which set it apart from simpler compression systems by supporting selfawareness: first, it is embodied, and second, it compresses observations of its own behaviour.
Compression enables the binding of information that is
distributed in both time and space. The embodied approach
in cognitive science emphasises the vital role an organism’s
body plays in determining how and what it thinks. According
to
The second ingredient concerns the social aspect of
consciousness. According to
We propose that this ‘understanding of the self’ is a requirement for the accurate modelling of the behaviour of others. If an individual lived in complete isolation within a simple environment, there would be no motivation for maintaining a complex model of a self that has experiences. It is only when people are embedded in a complex social environment that the goal of interacting with others requires them to maintain a detailed model of their own actions. When you observe others, they observe you observing them observing you. This recursive modelling, achieved through data compression, gives rise to a rich understanding of selfhood, an understanding of what it feels like to be conscious.
This recursive self-modelling may also explain another aspect of consciousness that seems to evade logical explanation, namely its subjective flavour, or qualia. Some examples of qualia include the pain of a toothache, the taste of sweetness or the perceived redness of an apple. These kinds of experiences seem to defy objective, reducible description.
We propose that the unique flavour of consciousness results from individual differences in how stimuli are compressed. Conscious perception is not a passive process whereby incoming sensory information is simply recorded in its raw format. Instead, people gaze through the lens of understanding provided by data compression. Different people, with different memories, will extract different patterns in a given stimulus, leading to different subjective experiences. In particular, people will have an intimate personal understanding of how they themselves experience or ‘compress’ a stimulus, one which remains beyond the remit of objective science.
In support of this position, Chalmers entertains the possibility of philosophical zombies, entities which act just like a conscious individual, but which lack qualitative experience. He argues that because such zombies seem conceivable, they must be logically possible, hence raising questions about the soundness of physicalism. Compressionism, on the other hand, takes the opposite view, seeking to bridge the explanatory gap. If consciousness is equivalent to data compression then any two systems which carry out the same compression should be viewed as equally conscious.
Chalmers’ (1995) acceptance of philosophical zombies paves the way for a double-aspect theory of information, leading him to speculate on the possibility of alien ‘psychophysical laws’ that govern qualia, supporting widespread ‘panprotopsychic’ consciousness across the universe. Compressionism, in contrast, offers a more parsimonious account of the link between consciousness and the physical world, relying on computational complexity rather than supernatural phenomena. By explaining how straight-forward information processing can bind information and lead to subjective, selfaware experiences, it might seem that compressionism can dissolve the hard problem.
Nevertheless, proponents of compressionism should not be
so quick to dismiss Chalmers’ views. For a start, optimal data
compression is not computable
Converging sources of evidence suggest a fundamental explanatory role for data compression in the evolved function of the brain. We have explored some of this research and sought to connect it together by proposing the term ‘compressionism’. The concept of data compression is already pervasive in the history of psychology and cognitive science. From neural networks, to classical ACT-R and SOAR architectures, to Bayesian predictive models, information is processed in ways that end up compressing events spread out over space and time. Thus, what compressionism offers is not a radically new idea but rather a framework for reconciling these varied approaches, one which opens up the use of powerful mathematical tools developed in AIT by Solomonoff, Chaitin, Hutter and others. Although it cannot and does not seek to eliminate the hard problem, compressionism can serve to put a name and structure to that hardness.