Hello everyone, my name is Adam Safran. I am a postdoctoral research fellow at Indiana University and today I will describe a synthetic theory of consciousness I recently proposed and published in Frontiers in Artificial Intelligence. It's called Integrated World Modeling Theory or IWMT for short and it's a somewhat complicated theory.
But I'll do the best I can to provide a high-level overview. For more details, I refer you to the publication, which is free to read online. I also have a companion preprint, Integrated World Modeling Theory Revisited, where I go into additional issues that I did not have space to address in the main publication.
So the goal of Integrated World Modeling Theory is to make headway on the enduring problems of consciousness, all of them. And towards this end, what I do is I cross-reference many of the leading models of consciousness, and there's many of them. And I specifically focus on integrated information theory and global neuronal workspace theory. And what I do is I try to look at what are the areas of overlap between these theories?
What are the areas of convergence? How are they similar? While at the same time, noting the differences and considering what... those differences might imply so in this integration attempt the way in which i bring these theories together is within an overarching framework of the free energy principle and active inference as pioneered by carl friston and colleagues which is gaining increasing popularity and controversy as potentially the first unified paradigm for understanding mind and life In a very brief nutshell, the free energy principle states that persisting systems must be doing something somewhat intelligent in order to persist in a world governed by the second law of thermodynamics or a world in which you would expect all things to get all mixed up and evolve towards states of maximal disorder.
So how is it that this maximally probable outcome is avoided? The free energy principle states that persisting systems in some sense must entail predictive models within their dynamics that helps to regulate these dynamics and exchanges with the environment so that the systems can hold themselves together, self-generate, regenerate, and not dissipate. The free energy idea is an information-theoretic-objective function. which you use to characterize how good are these models.
Within the free energy principle, there's a more specific process theory of the active inference framework, or what must systems do in order to minimize their free energy or their prediction error. There's even more specific ideas related to predictive processing or predictive coding in the Bayesian brain. And in the paper, I try to show how all of these things hang together and I try to give an example. explanation and review of these things. With respect to the Bayesian brain, in brief the idea is that perception is a kind of inference to the best guess as to the causes of sensory observations so that your perceptual experience is actually a kind of prediction or probabilistic inference of what you think is in the world.
To justify this idea, One part of it would be that your sensation is actually fairly impoverished and ambiguous. You only have visual acuity of about a thumbnail held out at arm's length. Further, even within that region, a good chunk of it is taken up by the blind spot.
Yet, that's not how your perception seems. There seems to be some sort of additional filling in process, this generation of additional information. that's more complete than you would expect given your sensations. Now the exact extent of how rich this is and how much fillion occurs is a matter of some debate, but the idea is that some kind of additional abduction or probabilistic inference is allowing you to infer a more complete sensorium from your sense data. This idea is gaining increasing traction in light of the Rise of deep learning and the success that it's been encountering recently in artificial intelligence so in the past people would focus more on discriminative models that are trained in a supervised fashion where you show them a bunch of stimuli you tell them what the stimuli are and then these systems are capable of telling you a hypothesis of What different stimuli constitutes so I?
train up these networks on a bunch of Google images, and then I can ask it, am I looking at a dog or a cat or a face or what kind of face? Now, with respect to consciousness, the more recent developments in deep learning involve an inversion of these generative models or the opposite direction of inference. Rather, instead of going from data to a hypothesis of what you're seeing, you go from your hypotheses or your models of what you expect and then generate likely patterns of data. So on the right there you can see the output of an early generation generative adversarial network producing images of faces that don't exist in reality but seem likely based on the training data.
And more recent incarnations of these technologies, it's really amazing what they can achieve. So amazing in fact that these form the basis for these deep fake images where you can't even tell the difference between whether this was a real or a fake image. So this type of mapping, this type of computational principle, the relationship between a probabilistic network or a neural network and the entailed probability distributions or models of these networks, this mapping I propose and others that is a bridging principle that we can use to go between brain and mind. But with respect to consciousness, rather than just the filling in of, let's say, visual images like these faces, what you would be doing is your brain functioning as a probabilistic generative model would be filling in your entire sensorium, your sight, your sound, your touch, and this filling in. Over time of likely sense data for everything you experience would be the processes that give rise to your consciousness as a kind of inference or prediction.
This idea has been discussed as your consciousness constituting a kind of grounded hallucination or a waking dream or a kind of fully immersive virtual reality. And I'd say that's all right. And that's the basic idea.
But with integrated world modeling theory, I try to make some more specific claims. I specifically draw upon this distinction as Ned Block proposed between phenomenal consciousness and access consciousness. Specifically, I take phenomenal consciousness to be experience, what it is like or what it feels like, or subjectivity or a point of view in the world.
I consider access consciousness or a consciousness to be awareness, knowledge of your experience, or the manipulability and reportability of the information of your experience. And this distinction, I think, is important because with respect to these different models of consciousness, oftentimes I believe they're talking past each other because they're using terminology different. They're referring.
to different aspects of consciousness or different kinds of phenomena related to consciousness. So with integrated world modeling theory the basic claims are this your phenomenal consciousness is what it is like to be the functioning of a probabilistic generative model for the sensorium of an embodied embedded agent. Access consciousness and possibly also phenomenal consciousness requires this information be integrated into a world model with spatial, temporal, and causal coherence for a system and its relationships with its environment. Access consciousness and possibly also phenomenal consciousness further requires self-models with autonomy or agency and the ability to engage in counterfactual or imaginative processing. That's integrated world modeling theory in a nutshell.
and the paper goes into far greater detail. In this synthetic model I draw upon integrated information theory and global neuronal workspace theory. With respect to IIT, it can be hard to explain but in some ways it's fundamentally simple.
IIT starts from phenomenology or the nature of experience and tries to identify necessary essential features for all experiences. So your conscious experience has intrinsic existence or it has a perspective, a particular perspective that's from within. It has composition, or there's different things within your consciousness. It has information and that your consciousness being one way distinguishes it from all the other ways it could be.
It's integrated and that your experience as a whole is greater than the sum of its parts. And if I were to remove one of the parts, that would be a difference that makes a difference and you would have a different experience. It's also exclusive in that I am conscious of some things and not of others.
And these phenomenal axioms are further stipulated to, or their mechanisms are postulated that might be able to realize these properties. And then what integrated information theory does is it tries to characterize different systems by their ability to realize these mechanisms, and then tries to characterize their potential for being sources of consciousness. IIT also has generated a lot of controversy.
Specifically, if those conditions, those axioms, are considered to be jointly sufficient for consciousness, then some of the entailments are things like quasi-panpsychism. IIT theorists have argued that things like 2D grids of logic gates, even if they're not referring to anything, could be highly conscious. Or you might have...
minimal consciousness associated with a single photodiode or an elementary particle. I believe that by treating the axioms as potentially necessary but not sufficient for consciousness, you end up sidestepping much of this controversy. And in terms of what would allow us to have sufficiency, I argue you need coherence. You need that these IIT analyses must apply to systems.
capable of generating world models with coherence with respect to space, time, and cause for the system and its relationships with the world. And so a system could in theory have an arbitrarily large amount of integrated information, but there still might not be anything that it's like to be such a system unless it was given this kind of embodied grounding that it could bring forth an integrated world model. So... In addition to integrated information theory, I also draw upon global neuronal workspace theory, which is quite different from IIT in terms of rather than starting from axioms of experience, it starts from what are the functional or computational properties of consciousness.
GNWT argues that consciousness involves the global access and broadcasting of information via these ignition events or these synchronous complexes where information from otherwise Isolated specialist processes or modules are, by becoming taken up into these large-scale reentrance signaling complexes, you get the formation of a temporary global workspace where the information can be made globally available and achieve what Dennett calls fame in the brain. And in more recent versions of GNWT, these ideas have been explicitly described. in terms of Bayesian or probabilistic inference where these ignition events where workspaces form are thought of as basically the selecting of a winning interpretation of what's going on in the world, of a winning model relative to a range of alternatives. And I think this is basically right.
And so I try to bring together IIT and GNWT as both capturing essential properties of consciousness but in different ways. And so IIT and GNWT, they're currently in some ways like at each other's throats with debates and there's this adversarial collaboration starting up to see which is the good theory. I think this is good that this is happening although I think Part of it, what we might expect, is rather than one theory being right and the other theory being wrong, it might be that they're both right, but they're right about different things.
Specifically, integrated information theory. argues that consciousness, the physical substrates of it, are located with a posterior hot zone at the back of the brain, while GNWT argues that conscious access involves frontal lobes, needs the frontal lobes to be involved in order to achieve this global availability of information and the ability to have knowledge of it and report on it. I think these are both correct statements and they're talking about different things. IIT, I'm argue in the manuscript is talking about phenomenal consciousness and I agree that it is likely primarily generated in the back of the brain, specifically posterior medial cortices, in particular the precuneus as the mind's eye or the Cartesian theater, if you will, and the inferior parietal lobule, specifically the angular gyrus and supramarschinal gyrus, as a source of body schemas or your body, the generation of body sensations as felt in all of its different aspects, and also your attentional or intentional awareness state. And this has relevance to attention schema theory.
And so if you bring together your attention schemas, your body schemas, and this visual model of the world, and you have them all integrated, I argue that this is what generates your consciousness. This is basic phenomenal consciousness. Although I also argue that GNWT is correct with respect to the frontal lobes being important for access consciousness or to have higher order knowledge of this experience.
Yet the experience itself in line with IAT I argue is always generated in the back of the brain. So it'll be interesting to see what the how the adversarial collaboration unfolds. and whether we find that one theory is right or one theory is wrong or that they're both right in different ways, which is what I claim. Integrated world modeling theory is ambitious in its scope.
I try to show how the free energy principle and integrated information theory can be combined and how both of these can be combined with global neuronal workspace theory to explain consciousness. I go into further details about computational principles from machine learning architectures that might be relevant for understanding how it is that the brain functions as a generative model that could give rise to experience. I also focus on relevant neural systems and in particular talk about ways that neural synchrony particularly at alpha frequencies or 8 to 12 times a second might be integrating information across your entire sensorium to create an integrated world model from a coherent egocentric perspective providing a point of view on the world with respect to the role of synchrony i address or draw upon cell and adasoy's connectome harmonics framework and i discuss some additional theories of consciousness and the potential functional significance of consciousness So with integrated world modeling theory, I make a fairly bold claim, which is that the hard problem might not be nearly as hard as we thought. Rather, we just needed to wait to have the right bridging principles as made available by advances in probability theory and machine learning to help us understand the types of computational processes that would be involved in the generation of consciousness. And we also needed to remember or remind ourselves the inherently embodied nature of conscious experience that what we are conscious of is Experience and our experience is fundamentally embodied.
So I'm moving forward. I'm going to more details on the ways that different neural systems contribute to different aspects of consciousness including with respect to intentional goal oriented behavior and That's what I've attempted to do and I'm very curious to know what people think. So if any of you read the paper and have any questions or opinions, please feel free to contact me. I look forward to talking to you. Thank you.