7. Information Theories of Consciousness

[…] to the extent that a mechanism is capable of generating integrated information, no matter whether it is organic or not, whether it is built of neurons or of silicon chips, and independent of its ability to report, it will have consciousness.

Giulio Tononi, Consciousness as
Integrated Information: A Provisional Manifesto1

Information is notorious for coming in many forms and having many meanings. It can be associated with several explanations, depending on the perspective adopted and the requirements and desiderata one has in mind.

Luciano Floridi, Information: A Very Short Introduction2

7.1 What Is Information?

In this ‘information age’ people see information everywhere. Some say that we are living in a simulation or a digital universe; others claim that information patterns are consciousness.

I open up your head and rummage around inside. I feel bones, blood and tapeworm cysts. Through the microscope I observe neurons, glia and bacteria. I cannot see information anywhere. I cannot detect it using scientific instruments. There is just soggy oozing physical stuff.

Computers are information processors. There must be information inside a computer. I open up a computer and rummage around inside. I feel silicon chips, copper circuits, dust and two dead flies. Just more physical stuff—no information anywhere.

I flick through the computer manual. It states that information is stored in the memory units of the computer (the DRAM storage cells). I switch on the computer and examine the DRAM storage cells. They contain electrons. When I measure the voltages I obtain the following values: 0.7, 0.8, 1.1, 1.0, 0.2, 0.7, 0.9, 0.1, 0.0, 1.5, 1.4, 0.5, 0.1, 1.5, 0.7, 0.8, 0.3, 1.2, 1.3, 0.0, 0.4, 0.9, 0.7 and 0.6. These voltages change all the time as the computer operates. If an engineer looked these voltages, s/he would note that the DRAM is operating in its specified range.

I apply a threshold of 0.75 V to the voltages, and interpret voltages above the threshold as 1 and voltages below the threshold as 0. This yields 011100100110010101100100. This means something to me—it is a string of 1s and 0s. A computer scientist exclaims, ‘Ah, binary, that’s 726564 in hexadecimal.’ A child interprets it as an adder.

I group the 1s and 0s into three 8-bit binary numbers: 01110010, 01100101 and 01100100. These correspond to the decimal numbers 114, 101 and 100. I map the decimal numbers onto letters using the standard ASCII codes (114=’r’; 101=’e’; 100=’d’). This yields ‘red’. It is a word in the English language (the colour of apples; the colour of blood). ‘red’ does not mean much to people who do not speak English—it is just a string of letters, similar to ‘nob’.

Initially the computer was an invisible physical object. I did not attribute any properties to it. It was something beyond my bubble of experience that did not exist for me. Its physical states were not 1s and 0s; they were not letters or numbers; they were not even voltages.3

Voltages, binary numbers and ‘red’ are information patterns that appear when we measure a system’s states and interpret the measurements in different ways. This combination of measurement and interpretation will be referred to as an interface, which specifies:

• The material that holds the information. In a computer the information could be in the DRAM, CPU, etc.
• The type of information. Is it binary, decimal, drawn from the set of letters, and so on?
• How information of the appropriate type can be read from spatiotemporal patterns in the material. In the computer example I specified how the DRAM voltages could be measured and converted into binary numbers and letters.

Interfaces enable us to extract information from the invisible physical world. They can be applied in sequence to extract different kinds of information. There is no information without an interface.4

An infinite number of different interfaces can be applied to a physical system. Instead of a threshold of 0.75 V I could have used a threshold of 0.55 V. This would have yielded 111101100110011101100111. I can group these 1s and 0s into four 6-bit binary numbers: 111101, 100110, 011101 and 100111, which correspond to the decimal numbers 61, 38, 29 and 39. I can use a different mapping of numbers onto letters (for example, 61=’b’, 38=’l’, 29=’u’ and 39=’e’). This interface extracts ‘blue’ from the voltages in the computer’s memory.

‘Red’ and ‘blue’ appear when I apply different interfaces to the DRAM voltages. There is no correct answer about which sequence of letters is really in the computer’s memory. Different interfaces produce different sets of information.

Once I have selected an interface, the information is determined by the physical system. If I interpret the DRAM voltages using a threshold of 0.75 V, 8-bit numbers and standard ASCII codes, I inevitably end up with the word ‘red’—I cannot change the fact that the application of this interface to this system results in the word ‘red’. While information can only appear through a subjectively chosen interface, it is fixed by the physical system once the interface has been selected—it is objectively present on the basis of this interface.

Custom interfaces can be designed to read most and possibly all information patterns from a physical system in a particular state. I can extract the text of Madame Bovary from the lines on my wife’s face.5 Think of a four letter word—I can extract it from the DRAM voltages by changing the number-to-letter mappings. Time-indexed interfaces might be required to extract complex information from simple systems6 and to extract sequences of information patterns from sequences of physical states.7

Some people distinguish data from information. They define data as the differences that are extracted from a physical system using an interface. These differences become information when they are well-formed and meaningful.8 The problem with this distinction is that any measured set of differences is meaningful to some extent: Voltages are meaningful to engineers; binary numbers are meaningful to computer scientists; letters are meaningful to literate people. The only differences that are completely without meaning cannot be accessed by us because they are part of the invisible physical world. This leaves us with the notion that information might be well-formed data. But we do not need a data/information distinction to capture the difference between well-formed and badly-formed data.

Shannon’s work on the transmission of information has led some people to interpret information as the reduction of uncertainty.9 Consider my snake, Sam. Sam is dead. Sam is not Lazarus: he will not rise—he will always be dead. You do not need to tell me that Sam is dead because I know that he is dead and this is not going to change. I do not gain any information when you send me a message stating that Sam is dead. Now consider a coin that can be in two states (heads and tails). I gain information (I reduce uncertainty) if you tell me that it is tails because I can only guess this with 50% accuracy. Now consider a six-sided dice. You roll the dice and it shows a two. I can only guess that it is showing two with 17% accuracy, so a message informing me that it is two considerably reduces my uncertainty about it. The more a message reduces my uncertainty about the state of a system, the greater the information content of that message. Shannon used this interpretation of information to develop his measure of information entropy.

This interpretation of information is a useful way of quantifying the amount of information in a system. But it is not an adequate definition of information. Before we can talk about the reduction of uncertainty of our knowledge about a system, we need an interface that defines the information states that are available in the system. We can only reduce uncertainty about the state of a coin once we have an interface that converts the physical coin into two possible outcomes, ‘heads’ and ‘tails’. Once a system’s information states have been defined, it is possible to measure its information entropy and state how rapidly its information can be passed over a channel.

7.2 Information C-Theories

Information c-theories are defined as follows:

D13. An information c-theory links consciousness to spatiotemporal information patterns. Information CC sets only contain information patterns, which can occur in any material.

Suppose we discover a neuron firing pattern, p10, that is correlated with conscious state c4. We could apply an interface, i1, to this pattern to extract an information pattern, ip1. An information c-theory would claim that c4 is correlated with ip1. This c-theory would predict that c4 would be present if ip1 was extracted from a pile of stones or from a set of traffic lights (see Figure 7.1).

Figure 7.1. Information c-theory. An experiment demonstrates that conscious state c4 is correlated with neuron firing pattern p10. Interface i1 converts neuron firing pattern p10 into information pattern ip1. An information c-theory would claim that it is the information pattern, ip1, that is linked to c4, not the neuron firing pattern p10. Information pattern ip1 can also be extracted from traffic lights through interface i2. The information c-theory would claim that ip1 is linked to c4 regardless of whether it has been extracted from a neuron firing pattern or a set of traffic lights. Image © David Gamez, CC BY 4.0.

Tononi has developed an impressive information c-theory. His algorithm analyzes the information patterns in a system, and outputs the parts that are linked to conscious states, the level of consciousness and a high-dimensional mathematical structure that is intended to correspond to the contents of consciousness. Preliminary experiments have been carried out to test this c-theory.10

Physical c-theories use interfaces to gather information about the physical world. The interface acts as a window onto the materials, and physical c-theories link patterns in these materials to consciousness. In an information c-theory the information pattern that is extracted through an interface is not a measurement of something else—it is linked to consciousness independently of the interface or the material in which it occurs.

Information c-theories can be converted into physical c-theories by adding material(s) to the CC sets.11 Physical c-theories can be converted into information c-theories by removing the material(s). The experiments that support Tononi’s information c-theory can be interpreted as evidence for a link between neuron firing patterns (identified using his algorithm) and consciousness.

Information c-theories are a radical departure from standard scientific practice. Scientific laws apply to specific aspects of the physical world. It is not the pattern that counts, but the presence of the pattern in a particular material. Newton’s theory of gravity describes how masses behave on a particular spatiotemporal scale. His equations would produce incorrect results if they were applied to electric charges. Information c-theories break free from the material—they treat information patterns as if they had an objective existence of their own—as if they were something in the physical world that could be linked to consciousness.

7.3 The Subjectivity of Information

A brain is in state s1; conscious state c5 is present. My laboratory carries out a pilot study to identify the information pattern that is correlated with c5. Tony chooses one interface and claims that the resulting information pattern, ip2, is correlated with c5. George chooses a different interface and claims that the resulting information pattern, ip3, is correlated with c5. Which information pattern is correlated with c5—ip2, ip3 or both?

Tony is my pal. George broke my microscope. I want to accept Tony’s claim that ip2 is correlated with c5. I want to reject the information pattern gathered by clumsy George. But the selection of ip2 would be an arbitrary subjective choice. If I want to conform to constraint C1, I have to accept that any and potentially all of the information patterns that can be extracted from the brain in state s1 are potentially correlated with c5. To avoid subjectivity my pilot study will have to measure them all. This is impossible because there is an infinite number of them.12

Suppose I use all possible interfaces to measure all of the information patterns that can be extracted from the brain in state s1. I now need to identify the ones that are correlated with c5. Which of these information patterns are not present in the unconscious brain?

My pilot study will have to use all possible interfaces to measure all possible information patterns in the unconscious brain. I can then compare these infinite sets to find the information patterns that are only present in the conscious brain. These are the members of the information CC set that is correlated with c5. The practical impossibility of this task suggests that a subjective choice of interface cannot be avoided in real world experiments on information c-theories.

These practical difficulties are irrelevant if interfaces can be custom designed to extract arbitrary information patterns from the brain (see Section 7.1). This would enable any information pattern to be read from the unconscious brain, including the information patterns that were extracted from the conscious brain in the first stage of the experiment. Custom designed interfaces that can extract arbitrary information patterns would break constraint C3. CC sets cannot consist of information patterns if all of the conscious brain’s information patterns can be extracted from the unconscious brain.13

7.4 E-Causal Powers of Information

I define an interface that interprets voltages in a computer’s memory as 1 if they are above 0.75 V, and as 0 if they are below 0.75 V. The information changes as the voltages change. This interface makes no difference to the patterns of e-causation in the computer—with and without the interface the computer moves through the same sequence of physical states.14

I alter the interface and specify that voltages above 0.8 V should be interpreted as 0, and voltages below 0.8 V should be interpreted as 1. Now the information patterns are completely different, but the computer continues to move through the same sequence of physical states. It does not matter which interface I apply to the computer: its e-causal exchanges and sequence of physical states remain the same. The information does not e-cause or constrain the behaviour of the physical system. This suggests that information cannot e-cause c-reports, so information patterns cannot be sole members of CC sets (C4).15

7.5 Is Information Intrinsic?

Information appears when an interface, defined by an observer, is applied to a physical system. Information patterns that depend on an external interface cannot be intrinsic properties.

It is conceivable that the interface could be inside the system, so that one part reads information from another.16 In this case the information might be an intrinsic property of the system as a whole. There are problems with this proposal. For example, the location of the information patterns in the system would be ambiguous, and information c-theorists have not proposed how we can measure this type of ‘intrinsic’ information without applying an external interface to the system.

7.6 Separating Information from Material

Suppose we identify a neuron firing pattern that is correlated with a conscious state. There are two interpretations of this result:

• A pattern of information is linked to the conscious state (information c-theory).
• A pattern in a material (neurons) is linked to the conscious state (physical c-theory).

We want an experiment that can decide between these two claims. This would show that an information pattern is correlated with consciousness (information c-theory), or that the pattern is only correlated with consciousness when it occurs in biological neurons (physical c-theory).

The best way of deciding between these claims would be to change the brain’s materials while preserving its information patterns. If it had the same conscious state when its neurons were replaced with silicon, then the information pattern might be the sole member of the CC set. But if we exchange a person’s neurons for silicon, we cannot be confident that their c-reports are functionally connected to their conscious states. We will have lost our ability to measure consciousness (see Section 5.4).

We have to use natural experiments to decide between the two claims. We could monitor the system and hope that the pattern moves between materials during its normal behaviour. Suppose the subject has conscious state c6 when there is information pattern ip4 in the neurons and nowhere else in the brain. At a later point in time the subject has c6 when ip4 is in the glia and nowhere else in the brain. We would conclude that c6 is correlated with the information pattern, and that the material has no effect on consciousness.17

We have no reason to believe that information patterns move between materials during natural experiments on the brain. If we cannot observe this, it will be impossible to experimentally distinguish between physical and information c-theories.

7.7 Summary

Information appears when interfaces are applied to the physical world. Interfaces specify how information of a particular type can be extracted from a particular material. Information does not exist in the physical world—it is partly determined by the interface and partly determined by the physical world. Information c-theories claim that information patterns are linked to consciousness independently of the material in which they occur.

Information patterns cannot be correlated with consciousness because they can be read from both the conscious and unconscious brain using custom-designed interfaces (C3). Information patterns are subjective and incapable of e-causing c-reports (C1, C4). It is unlikely that evidence in favour of them can be obtained through natural experiments. Until these problems have been resolved information c-theories should be set aside or interpreted as physical c-theories.18