Towards a representationalist theory of (machine) Consciousness
Part 1: Phenomenological Axioms
Integrated Information Theory (IIT) starts from five axioms, aimed to capture the essential aspects of every possible experience. These five axioms are then used to construct an information-theoretic theory of consciousness based on the physical cause-effect properties of a system. I believe this is an excellent approach to creating a theory of consciousness, except that the level of explanation, the level of cause-effect properties of physical systems, is too far removed from the original phenomenon the theory is about.
In this series I propose to explore what happens when the axioms are translated to a level much closer to the original starting point, the representational level. Specifically, I propose a translation to the properties of models, with the aim being that the theory is equally applicable to biological and artificial systems.
This first instalment looks at the axioms in detail and suggests some possible updates.
Properties of Experience
Ernst Mach's 'self portrait' has been used (including in IIT) as a depiction of subjective experience. Seen as a self portrait, it nicely interprets the self as being the content of our unfolding experiences of the world (a view of the self I have briefly talked about previously). It coincides with the idea that we are our best model of our inputs. Though, in reality, the constant, uninformative, components (such as our nose or exquisite moustache) would be filtered out before making it to consciousness.
As this blog is focussed on more than just human experience, and especially that which could arise in AI, perhaps a more appropriate version, would be the self portrait of the DOOM playing AIs from Ha and Schmidhuber's World Models.
for the original IIT versions.
There are two claims in this axiom.
- Consciousness exists (it is real).
- Consciousness exists from its own intrinsic perspective.
The second claim, that experiences are subjective, is one of the defining features of consciousness. The famous description of consciousness of x as meaning 'there is something it is like to be x' shows how subjectivity is at the core of the phenomenon. The thing that it is like to be is the subject of the experience.
The first claim is stronger and comes with a lot of ontological baggage. Consciousness must seem to exist to the entity that is the subject of it, but it is certainly debatable how much more can be said. The debate has included eliminative materialists, instrumentalists, illusionists, mild realists, and realists. As the future derivations require nothing more than a seeming existence, then I propose this is all that needs to be said.
Update: Replace with
- Consciousness seems to exist, and from its own intrinsic perspective.
Consciousness is structured. There are multiple components (phenomenological distinctions) within each experience.
This axiom seems to skip a step. If we are starting from the most fundamental properties of experience, such as its subjectivity or seeming existence, then the fact that experiences even have content should be explicitly stated. This means that the theory is committed to saying that a truly contentless experience cannot exist. This doesn't mean that an experience cannot be an experience of nothingness, because then nothingness would be the content.
Update: add axiom Intentionality
Experiences have content. They are of something.
The claim makes sense in all experiences we are familiar with, and is more an axiom about the possible nature of experiences than its necessary qualities. To ensure that it is a necessary condition it can be weakened to include the possibility of a single component.
Interpret multiple as one or more.
The phenomenological content of each experience differentiates it from all other possible experiences.
This is another way of saying that experiences are defined by their contents (in other words we can tell different experiences apart when they have different contents). This was introduced above as the intentionality principle. In the IIT formulation it seems to also require that the contents of experiences are commensurable across different subjects, which could be true at the informational level that they translate to, but not at the descriptive level that the axioms are introduced by. For clarity it seems best to fold this claim into the intentionality claim introduced above.
Update: folded into Intentionality above
Consciousness is unified. Each experience is irreducible to the separate components that it may contain.
This is a very important axiom for the theory and one that also, in my opinion, provides the deepest insight into consciousness without going beyond axiomatic status. It is the subjective nature of experience that binds the content into one unified whole.
However, as written the axiom does not make a distinction between integration of content and integration over time. Another way of writing the integration of content is to say that a subject only has one (integrated) experience at a time. However, experiences can extend over time, potentially changing their content. The constant unifying factor of a changing temporally extended experience is the subject of the experience.
Update: split into Content Integration and Temporal Integration
- Contents of consciousness are experienced unified at any given time.
- Experiences can change over time and appear unified to the subject.
Consciousness is definite in content and spatio-temporal grain.
The uncharitable interpretations of this axiom do not look good. In the original it states that "each experience has the set of phenomenal distinctions it has, neither less (a subset) nor more (a superset), and it flows at the speed it flows, neither faster nor slower." This is either an empty claim (even for an axiom) or a strong claim about the possibility to give definite demarcations to all experiences by both content, spatial properties and temporal properties. It seems far too strong to claim that the length of time an experience lasts has an objective measure, when this is only experienced subjectively and varies wildly with external observation. I have also previously discussed the indefinite nature of the content of our experiences.
A more charitable interpretation of this axiom is that it is about the fact that the content of our experiences always seem to be definite in the sense that we experience a single possibility out of all the possible configurations of the potential contents of our experience, as opposed to experiencing all the possibilities at once. However, our experiences are not totally definite. If I take a playing card and look at it only in my peripheral vision (without knowing which card it is), it will not be clear what I am seeing. I don't experience exactly one of the 52 possibilities, but something indeterminate, shaped by the possibilities, but not exactly bound to any single one of them. This is partly down to a lack of sensory precision, but axioms of all possible experiences should not depend on good sensory apparatus.
Update: drop (for now)
I will return to these later in the series and also consider extensions to cover other components of experience.