The Meta-management theory of Consciousness
The computational theory that explains the evolution of phenomenological consciousness
We have been trying to figure out what consciousness is for hundreds of years, probably more. Most researchers in the field today will probably tell you that we are still no closer to answering the question than when Descartes first suggested his dualistic theory. The problem is that there are so many conflicting ideas out there and there are strong arguments that challenge the very notion of being able to approach this question scientifically. I think we know more than we’re giving ourselves credit.
I’ve been developing a theory of consciousness for the last decade. It has similarities and ties with several other theories and draws inspiration from what we know about the neuroscience of the brain, and from what we’re learning about intelligence via the ever-growing field of machine learning. However my theory is unique in that it is rooted in the need to provide an answer to why consciousness evolved, and in the clarity of its answers.
In this post, I will present an accessible description of the Meta-management theory of consciousness, and I will show how it answers important philosophical and practical questions about consciousness. Rooted in evolutionary principles, it explains why consciousness evolved, giving us some hints about which other animals experience consciousness. It explains how consciousness and intelligence are deeply intertwined, giving us clues about general intelligence. It also encroaches onto the so-called “hard problem” of consciousness by explaining why we only have conscious access to certain aspects of cognitive processing, why it “looks through” to so-called first-order states, to what extent consciousness has causal influence over thought (and to what extent it doesn’t) and why consciousness feels the way it does.
· Context
· Theory in Brief
· Deliberation
· State Trajectories
· On Models and Planning
· Meta-management
· Capturing Cognitive State
· Auto-meta-management
∘ The internal critic
· Voila! Consciousness
∘ The cognitive self
· Consequences for Understanding Intelligence
∘ Evolution
∘ Abstract thought
· Consequences for Understanding Consciousness
∘ Limited access consciousness
∘ The intentionality of consciousness
∘ The causal nature of consciousness
∘ The functional purpose of consciousness
∘ Consciousness by degree
∘ What consciousness “feels like”
· Closing Remarks
· References
Context
Before I start, let me provide a little context to the way that I approach such a question.
Consciousness:
- Consciousness is the first-person subjective experience associated with being a living human being (maybe others too, but we’ll come to that). It encompasses having awareness of our own thoughts and perceptions and knowing that we are aware of those things (so called access consciousness). Consciousness also encompasses the way that access to (some of) our inner workings doesn’t just merely occur, but is associated with an inexplicable feeling — the so called “feels” of phenomenological consciousness.
- I am interested in why certain creatures (eg: humans) have conscious experience at all, and why some aspects of brain function are associated with consciousness and make up the contents of conscious experience, while other aspects do not. In other words, why some state is consciously accessible, while other state is not.
- I will be talking here about all of the above aspects of consciousness.
Materialism:
- I take a materialist view of the nature of consciousness. This means that I believe that consciousness can be explained entirely in relation to the known physical universe, without any recourse to metaphysics.
- Further to that, I strongly suspect that consciousness can be explained reductively — counter to many philosophical positions today.
- I also don’t see any problem with taking on a particular “belief” here. Almost all researchers in the field have clear and strongly held beliefs one way or another.
Phenomenalism:
- Unlike some materialists, I embrace the fact that consciousness feels phenomenal — it has an associated feeling that seems inexplicable in mechanistic terms. I 100% agree that any complete theory of consciousness must explain that feeling.
- I just think that there are materialistic explanations for the phenomena.
Design stance:
- My approach to understanding consciousness is to apply something called a “design stance”. This means that I’m trying to understand how one might go about building such a system. This becomes a litmus test for understanding — do we know enough to build it yet? If we built something based on our current understanding, how would that compare to human consciousness?
Exemplars:
- Human consciousness is our most intimately familiar example. While we have some clues and intuitions about the animal kingdom, we know very little that we can be certain of. Thus, when it comes to examples of consciousness, humanity is it.
- However, when it comes to brain function, we know enough about the differences across the animal kingdom that we can use animal brains for some discussion points about cognition more generally.
Lastly, before I start, I need to explain one piece of terminology to avoid getting caught up in knots. The brain is a physical thing, and whatever goes on within it happens due to its neural physical substrate. Yet, there are times when it’s convenient to make a distinction between brain activity and the activity of the rest of the body (and the environment around it). For the latter, I can’t find a better term than physical. So, for this article:
- physical = the stuff of the body and its environment, excluding anything that the brain is doing.
- cognitive = the stuff of the brain and its internal processes (aka the mind)
Theory in Brief
I’ll start by giving up most the spoilers and providing a summary of the major aspects of the theory. Hopefully, some readers will find it interesting enough to read further. The subsequent sections will go into the details of those major aspects, followed by discussion of the consequences for our understanding of human intelligence and consciousness. Those later sections provide clear answers to some of the more puzzling questions about phenomenological consciousness.
The theory goes like this:
- Simple brains immediately produce actions in response to sensory stimuli. Advanced brains use deliberation: extended periods of “thinking” over multiple cycles of cognitive processing.
- During deliberation, there is no action/effect feedback from the environment or body to the cognitive processes to keep the cognitive processes “on task”. Just as for body control and interaction with the environment, something needs to monitor and control the cognitive processes during deliberation. This is known as meta-management.
- It turns out that it’s possible for the cognitive processes to meta-manage themselves, something I call auto-meta-management. This is possible via two mechanisms. (1) A feedback loop that captures cognitive state and makes it available as a cognitive sense, which is treated just like any other exteroceptive or interoceptive sense. (2) Predictive processes that are capable of modeling arbitrary cause-effect relationships and using them to determine actions.
- The meta-management feedback loop has several constraints and evolutionary pressures applied against it. In general terms, it will only capture state from the smallest cross-section of cognitive processes and use the most coarse-grained representation that happens to be necessary to effectively support meta-management processes.
- Consciousness is the result of the cognitive processes having access to observe their own state while they carry out their processing. Many puzzling aspects of the phenomenology of consciousness can be explained by the fact that it is a result of the above mechanisms.
Such a description may appear too simple to capture all the nuance of human consciousness, but I will show that it explains human consciousness very well. For example, the Meta-management Theory of Consciousness provides clear explanations for why we only have conscious access to certain aspects of cognitive processing, why it “looks through” to so-called first-order states, to what extent consciousness has causal influence over thought (and to what extent it doesn’t) and why consciousness feels the way it does.
Now let’s get into the details, starting with what deliberation is and why it’s significant.
Deliberation
Humans think a lot. Sometimes for long times. We think to carefully consider our actions before carrying them out. We mull over the reasons why our partner said what they said. We mentally work through math problems. We ponder the basis of consciousness. Some animals probably do something similar, in their own way. We feel uncomfortable calling that thought, so we use a different term — deliberation.
Deliberation is an extended period of cognitive processing that doesn’t immediately influence physical action (remember that “physical” is defined in this article to refer to the body and environment but to exclude anything to do with the brain). All mammals contain a well-developed cortex and are capable of modeling the world around them. They use that to simulate different possible courses of action before choosing what action to take in a given situation. There’s evidence that some invertebrates can do planning too. That’s deliberation.
There’s a good reason for brains to use deliberation. Brains are extremely energy-intensive. Evolution favors smaller, more efficient brains. For any given brain, there is a limit to what can be achieved by a single pass through its processes. Allow the brain to perform multiple passes, iteratively improving upon its previous conclusion each time, then the same brain can do much more.
There’s also a good reason why deliberation isn’t something we use much in ML today. It’s hard to control. Deliberation may occur with minimal to no feedback from the physical body or environment. Were that feedback available, it might provide important information about whether the deliberative process is leading toward a useful outcome or not. Without that feedback, how does the brain know whether it’s doing anything useful?
State Trajectories
To begin to answer that, it’s helpful to have a metaphor to work against. The metaphor that we’ll use here is that of state trajectories.
In the physical world, a state trajectory might describe the path taken by the individual as they navigate a maze, for example. It also might capture aspects of the individual’s body as it moves, such as the motion of its legs. Perhaps we might also consider the homeostatic processes that adjust heart rate, digestion, etc. to ensure that the body is functioning optimally according to the situation. For simplicity, we’ll refer to both generically as the individual’s physical trajectory.
Likewise, we can consider the state trajectory of the individual’s cognitive processes, which we shall call its cognitive trajectory. For example, the cognitive processes go through a series of distinctly different states as an individual sees something that they desire, makes the decision to obtain it, plans their course of action, and then carries out those actions. Importantly, during some periods within that sequence, there is close feedback between the individual’s cognitive and physical trajectories, while during the deliberative periods there may be no feedback. This is significant because both physical and cognitive state trajectories can get complex.
A state trajectory draws out a path through state space — the set of all possible states, combined with how they relate to each other. The illustration above gives example representations of a physical state space and a cognitive state space. In the former, a bird’s-eye-view of a maze captures the fact that there is a start position and a goal position, and walls that block motion. Additionally, there is an area of state space that needs to be avoided due to some inherent danger. The cognitive state space is represented as a contour map that captures the local minima and maxima associated with whatever search is being performed at the time.
It can be seen that the cognitive state space is just as complex as the maze, or even more so. And there are inherent dangers too. For example, Beaudoin (1994) has identified five common problems associated with systems with deliberative capabilities:
- Oscillation between decisions. Wasteful re-assessments of decision points, leading to a meta-stable (oscillating) but stagnant state (ultimately achieving nothing useful).
- Insistent goal disruption. Repeatedly getting distracted by competing goals that have been previously disregarded.
- High busyness. Attempting to multi-task between too many goals, leading to poor outcomes.
- Digressions. Choosing to deliberate over some sub-goal, and then losing track of the “big picture” by forgetting to return to the overarching goal.
- Maundering. Getting stuck deliberating over the details of a goal without making a decision.
The brain solves the problems associated with physical trajectories by learning, observing, modeling, and planning. It likewise needs to do the same for cognitive trajectories.
On Models and Planning
Let’s talk a little about those models before we get onto meta-management.
There’s growing evidence that the brain employs two strategies for the production of voluntary behavior, with any given action being the result of some combination of those two strategies (Douskos, 2018; Cushman & Morris, 2015; Dolan and Dayan, 2013).
The model-free strategy efficiently produces habitual (or automatized) behavior for oft-repeated situations. Internally, the brain learns something akin to a direct mapping from state to action: when in a particular state, just do this particular action. The model-based strategy works in reverse, by starting with a desired end-state and working out what action to take to get there.
In novel circumstances, the model-based strategy performs better than the model-free one. It does this by observing the world and its various interactions and building up a model of the world. This model approximately captures the separateness between different objects, and the cause-effect relationships for interactions with those objects and between those objects. With that model available, the brain is able to consider different possible actions and to predict their likely outcomes. Thus, when attempting to achieve an outcome that the individual has never attempted before, they can use the model to identify a suitable behavior.
As mentioned, the brain employs both strategies, trading off their relative strengths and weaknesses for different circumstances. For example, when navigating a maze environment for the first time, the individual may quickly build up a model of the maze as they walk around it, which they can then use to figure out how to navigate it for various different tasks. In contrast, if they are very familiar with the maze, they may navigate habitually instead — just like we do every time we absent-mindedly walk to the fridge. It is through that combination of habitual/model-free and model-based control that we take efficient paths to our targets while avoiding walls, other obstacles, and danger zones, and while we’re thinking about something else.
It’s worth also noting that the models we use are not just restricted to the domain of our external environment. We also model our own body: its components, how each component tends to behave, how they interact, as well as their current state. The planning activity we take when using the model-based strategy can incorporate models about both the environment and the body together. For example, when carefully treading across a floor covered in Lego pieces, we are using our model of the Lego pieces and the pain that they can inflict, and our model of the carpet and its relative slipperiness or grip, in conjunction with our model of our body, so that we can balance on one foot while carefully swinging the other leg around to find the next available space large enough to take the other foot.
Meta-management
There is a catch-22 problem that we are rapidly approaching as part of this narrative. We’ve identified that the physical trajectories carried out by the body are complex, and the solution is that the brain models, monitors and controls much of the body activity during those trajectories. In that description, there is a clear separation of responsibilities: the body does the action, the brain chooses what action to take and checks the result (actually, the separation is far more nuanced than that, but the simplification doesn’t impact our story). In contrast, when the brain state needs to be monitored and controlled, the only option is for more brain to do the monitoring and controlling.
I’ll return to that catch-22 problem shortly. But first, I need to introduce a little more terminology:
- First-order process: this refers to all of the cognitive processes that we’ve been talking about so far. They’re “first order” because they’re the main processes that take sensory signals about the physical world, consider them, and then produce physical action. A first-order process continually monitors the state and ongoing activity of the body, and it controls that state and behavior according to whatever needs are relevant at the time.
- Second-order process: this refers to a process whose task is to consider other processes rather than the body. It continually monitors the state and ongoing activity of other processes, and it controls the state and behavior of those processes according to relevant needs. Second-order processes will often be primarily focused on homeostatic needs for the proper functioning of the first-order processes.
Earlier, we identified that the cognitive state space is so complex that it too needs to be modeled, and cognitive trajectories through it need to be monitored and controlled. This is an example of a second-order process. It is known as meta-management.
Meta-management as a term isn’t used commonly. I take that as evidence that this approach to understanding consciousness has not received the attention it deserves. Researchers studying anything akin to what I’ve been talking about would more commonly use the term meta-cognition. But meta-cognition includes many different meta behaviors, from awareness of one’s abilities and choices of learning strategy in school to awareness of mistakes and the perceived accuracy of our memories. In contrast, I consider meta-management to be one of several low-level mechanisms underlying meta-cognition.
Capturing Cognitive State
Any process that monitors and controls something needs a mechanism to observe it. We can monitor and control our body because we observe its state through our exteroceptive (sight, sound, smell, touch, taste, proprioception, balance, etc.) and interoceptive senses (eg: heart rate, internal temperature, gustatory signals).
To monitor and control its cognitive processes, the brain needs to observe its own state. How should it do this?
Part of that state may be encoded within the internals of individual neurons. Other parts of the state may be encoded as recurrent firing patterns between groups of neurons. In the extreme, every single neuron and every single neural interaction contributes to the state of the cognitive processes. If we assume optimistically that each neuron in the meta-management system can capture and understand either the state of a single target neuron or a single target neural interaction, then a brain with 1 billion neurons and 1 trillion synapses needs an additional 1 billion + 1 trillion neurons for its meta-management. That sounds unrealistic at first, but all it really suggests is that 50% of the brain devoted to meta-management of the other 50%. However, the full state of the brain has just doubled. So to capture that state and to consider it requires a second doubling of brain size. This explodes in an exponential route to infinity.
Obviously, this is untenable, and perhaps a little fanciful. It’s also entirely unnecessary.
The knowledge we obtain through our senses about the world around us is informationally sparse, and yet we do very well with it. Correspondingly, we only need to capture a small amount of information about our cognitive state. The specifics of exactly how much and which state will be narrowed down through evolutionary pressures to the point where it is the minimum needed for effective meta-management.
Notice also how our conscious access to our various sensory modalities differs significantly by modality. We’ve known for a long time about our so-called “5 senses”. It’s only recently that the list has grown to include others such as the proprioceptive and vestibular senses, or for us to identify that the senses of heart rate and digestive signals are distinct from the sense of touch. Part of the reason for the delay is that we experience those other senses in either a much less clear way, or that they are simply much harder to become aware of.
Perhaps we have a clear conscious experience of the senses that are related to things that we consciously control. For example, we have clear conscious experience of the 5 senses because they directly inform us about actions that we might need to take. In contrast, the other senses that are less clear are perhaps associated more with automatic processes that we don’t have much conscious control over. Those statements are fairly circular, but I can suggest a more concrete variant. The distinction may lie in the balance between evolutionarily hard-wired vs learned processes, and whether model-based control is of use. In brief, I’d suggest that those behaviors that are learned and can be mediated by model-based control have associated senses with high acuity in our conscious experience, whereas senses involved with genetically hard-wired homeostatic processes will be heavily attenuated or omitted from conscious experience.
Here I am alluding to the way that we are consciously aware of our cognitive state and that it has relatively high acuity.
Auto-meta-management
Back to our catch-22 problem mentioned earlier. How might the brain do meta-management?
One option would be for the brain to evolve a separate region (let’s call it a second-order region) that does the monitoring, modeling, and controlling of the first-order cognitive processes (in what we’ll temporarily call the first-order region). That’s not as straightforward as it might sound at first. The second-order region needs access to all sorts of capabilities that are already present in the first-order region. For example, all those models. So now the brain needs to somehow wire up parts of the first-order region to be used by the second-order region. This sounds like some sort of complicated time-share system. Perhaps it’s feasible, but it turns out that there’s a simpler way.
Here’s a crazy idea: what if the first-order process could meta-manage itself? Something I call auto-meta-management.
It works like this… Optimized by evolutionary processes, the minimally effective set of first-order cognitive state is captured, producing a low-dimensional summary of that state. This is then made available as a sensory input to the first-order process. As per any other sense, the first-order process monitors and models the cognitive sensory input. That model captures the cause-effect relationships between different cognitive states, including identifying desired and undesired outcomes. The first-order process is thus able to predict the best actions according to the goal at hand and to use knowledge of its own cognitive processes to avoid unwanted cognitive behaviors.
There are several points worth drawing out:
- The cognitive sense is treated no differently than any other sense — it is a first-class sense. This ensures that all the same capabilities for modeling and for inferring latent state are available.
- Model-based and model-free/habitual behaviors work in conjunction here too. Through experience, the brain learns ways of thinking, including efficient strategies for problem-solving. These become habitually managed over time, saving the more computationally expensive model-based systems for other more interesting work.
- With predictive modeling processes available, there’s no difference in how the brain needs to respond to the environment, the body, or to its own cognitive state. In all three cases, it monitors and models them independently, cross-models them to understand the relationships between them, and then uses those models to decide the best course of action for a given goal.
At first, the idea of auto-meta-management may seem strange, or even outright impossible — it seemed unlikely to me too. But I think you will find that the idea grows on you. There is a simplicity and elegance in this proposal that avoids the pitfalls of other meta-management strategies. While it may seem that such a system could never stabilize, I believe it is possible with the right balance of evolutionarily hard-wired modeling and prediction behaviors plus learned cognitive processing, and with developmental stages adjusting which systems are able to take control as the brain develops.
More importantly, this suggestion fits very well with our first-person perception of consciousness. For example, finally, we have an explanation for why our thoughts are associated with conscious experience as well as our perceptions — because our thoughts are also available as a sense.
The internal critic
There’s a further precedent to support the argument presented so far. It has to do with the problem of sparse feedback.
It’s now generally accepted that the brain employs something akin to the Actor/Critic reinforcement learning approach used in ML (Bennet, 2023). According to this approach, an actor network learns to generate actions based on current state. Separately, a critic network learns to identify the inherent value of any given state. The critic provides detailed reward information that is used in the training of the actor. For example, if the agent needs to learn to navigate roads to get from a starting position to a finish position, the critic punishes the agent for going off the road. Additionally, if, let’s say, there are multiple possible routes to be taken, the critic provides rewards and punishments that encourage the agent to take the best route. This solves the sparse feedback problem in reinforcement learning: that actual rewards to the agent may only be granted at the finish line. If the actor network is trained on that feedback alone, it will take a very long time to train. The agent needs to internally generate its own signals, and the critic does just that.
The actor/critic approach has been used very successfully within ML to train self-driving cars, robots, and various other things. As mentioned earlier, it’s now believed that the brain employs it too as a guide when generating physical motion. This also makes sense for the monitoring and control of cognitive trajectories.
When a lengthy deliberative process is required to solve a particular problem, the only feedback available may be solely internally generated. A critic could learn the domains of cognitive problems and provide feedback during deliberation, estimating the value of being in a given state. This could even include the identification of common times when cognitive processes tend to get stuck or off track. Feedback from the critic would help train the agent, as per reinforcement learning.
There’s also no reason why that critic valuation information can’t be made directly available to the actor so that it can make on-demand changes when it realizes that it’s getting stuck or off track.
Voila! Consciousness
So here’s the punch-line: consciousness is the result of auto-meta-management via a meta-management feedback loop providing a cognitive sense that is inferred as conscious state.
The content of consciousness — whatever we happen to be consciously aware of — is a direct result of the state that is captured by the meta-management feedback loop and made available as sensory input. Thus we are aware of our thoughts because they are captured by the feedback loop so that we can meta-manage them. But it doesn’t stop there. While we consider our thoughts, that same feedback loop also captures the fact that we are now considering prior thoughts. Then we can consider the fact that we can consider thoughts. Putting that all together, combined with short-term memory and modeling, we construct a concept: that we are aware of our own thoughts.
Let’s put on our design-stance hats for a minute and run the litmus test. If we were to build the system described in some form of AI, what would it do? Let’s assume that it has sufficient modeling capacity to make sense of its environment, its body, and its cognitive system. Let’s also assume that it knows language and the culturally shared concept of consciousness. We’ll also assume that it contains an innate drive to seek knowledge, such as by the free energy principle. For simplicity, we’ll treat the cognitive system as if it executes by a simple feedforward process, and that deliberation occurs as multiple iterations of that process.
What information would the cognitive process have available to it while performing the mundane activity of observing an apple on a table, and what might it be able to conclude?
- The visual sensory signals capture the raw light pattern reflecting off the apple. Assuming the predictive theory of perception, internal models of the world infer the latent cause of that visual sense as being an “apple”. Those models don’t just capture the appearance of apples, but a whole cohort of other related information such as their weight and juiciness when squished. Perhaps even their taste and associated feelings of satisfaction from eating them, if this system were to eat.
- The recognition of this apple is open to deliberation by cognitive processes. Regardless of whether deliberation was to occur in this iteration or not, meta-management processes need to observe this moment of cognition and consider it in terms of the current context, such as goals and pressing concerns, so that deliberation can be encouraged or discouraged appropriately. Thus, even if only considering the apple for a moment, the meta-management feedback loop captures the fact that the cognitive system was considering the apple and makes it available as sensory input for the next iteration of cognition.
- As the cognitive sensory signals arrive on that next iteration, they are included along with other sensory modalities and partake in another iteration of perceptual prediction. Inferences are drawn about the latent cause of those sensory modalities, with attention modulating which sensory signals are given significance.
- We’ve already considered the inference of “apple” from the visual sensory signal at this moment, with all its detail and connotation. What of the cognitive sensory signal? We know that the true latent cause is the previous iteration of cognition. What might the system infer for itself? This will require a little more discussion to flesh out fully, which I’ll do below. In the meantime, let me just state simplistically that if the system can develop a concept of “apple” from its observations of exteroceptive senses, then it can also develop concepts of “body-self” and “cognitive-self” from its observations of exteroceptive and interoceptive senses (including the cognitive sense). Thus, perceptual prediction infers that the information obtained through the cognitive sense arose from the cognitive-self. It also associates modeled connotations to it, such as the concepts of “self”, “thought”, “desire”, and “volition”.
- Should the cognitive system attend to that cognitive sense, it may initiate a period of deliberation w.r.t. to it. Each iteration of that deliberation produces a new cognitive state that is made available in the next iteration for further consideration. Through the use of short-term memory or otherwise the change in cognitive state over time forms a trajectory that can also be considered. Thus every iteration processes a combination of three sources of information: i) the immediately prior cognitive state and its trajectory so far, ii) pre-existing knowledge about the system’s own cognitive processes that it has built up from observation over time, and iii) culturally shared knowledge about the concepts of self-awareness and consciousness.
- The system thus has all the information needed to conclude that it, too, is conscious. Not just the binary fact, but that its experience of consciousness carries many connotations. Every moment of awareness includes not just raw observational information but is associated with all the pre-existing knowledge and meaning that it has acquired over time. Its experience of consciousness has depth and nuance. It varies in form and acuity depending on attentional focus — whether the system attends to a particular visual-sense-inferred object or a cognitive-sense-inferred thought. The system experiences consciousness with a range of phenomenology akin to human experience.
What might be missing between this description and true human consciousness? I can think of nothing that would not be covered by the simple auto-meta-management mechanisms described, by cause-effect modeling, by predictive processing, and by the reasonable addition of other mechanistic human systems such as memory and emotional affect.
The cognitive self
I would love to leave the description here as it is and talk about some fascinating consequences for our understanding of consciousness and human intelligence. But I expect that some will find there to be a leap of logic in how the system reached conclusions about its own thoughts. So I’ll dwell on that for a little before proceeding. This discussion will be a little more technical than the rest and it doesn’t feed into the remaining sections, so feel free to skip it if it doesn’t interest you.
It’s clear that the system can infer latent state from its sensory inputs, that the inference depends on models of cause-effect relationships, and that it builds up those models from observation. We need to understand what those models might be in relation to the cognitive sense, and thus what it can infer from any given cognitive sensory state.
The problem for us is that we already know what caused the cognitive sense and we have strong expectations of what it should feel like from our own experience. It will help if we can prevent our own experience from obscuring the analysis. So let’s trick ourselves into thinking that we don’t know the true source of the cognitive sense but just calling it sense Z. For a baseline, we will add senses X and Y, which correspond to different exteroceptive modalities. We will also add motor modality M.
Remember that the system has powerful modeling capabilities, where models capture cause-effect relationships both within a given modality and between modalities. In the long period of learning, the system observes patterns within its sensory modalities and between them. For example, it observes countless cause-effect relationships between different states inferred from X. Through that, it learns to differentiate discrete high-level objects and to model cause-effect relationships between those objects. Thus it builds a latent model that represents not just raw sensory signals, but the actual structure of the world that results in (or generates) those sensory signals.
It also observes a different kind of cause-effect relation — that it as power to influence those objects. When it elicits actions via M, sometimes they influence X and Y states. How does the system identify that M was the cause of X and Y?.
In short, I’m not sure, but I can guess. There’s a few possibilities in how humans might do that. One is that there is some sort of direct observation of planned action and that is made available for modeling of cause-effect relationships. Another is that we build up such models through indirect observation. For example, that we somehow conclude that our body is our own, that our body contains arms, and that then when the arms move they do things against the environment. Another possibility is that our cognitive sense is what provides that observation, enabling us to build the cause-effect relationship from intention to action to result. For the sake of sticking to just the mechanisms described in earlier sections, I’m going to assume this last possibility.
There is a clear way that humans develop a sense of “body-self”. Our bodies are full of senses that respond immediately to touch and pain. Our visual sense enables us to see our arms, for example. Our proprioceptive sense enables us to know the position of our arms even when not looking at them. Together they feed into each other to develop a model of what arms are. As we use those arms we touch other parts of our bodies we identify cause-effect relationships between the arms touching the body and the perceptions that we have in return. This ultimately leads to us concluding that we inhabit physical bodies with a specific boundary. Thus we have a sense of “body”, but not necessarily a concept of “self”, nor of our own volition in those actions.
During that same process, we also receive cognitive sensory signals that capture intent. At first we don’t know of it as intent, but we do develop a cause-effect model that relates those pre-action states to subsequent action of say, the arms. Finally, as we observe ourselves versus others, and we observe the level of information we have about our own pre-action and action states versus that of others, we can develop a clear sense of “self” versus “other”. Thus now we have a concept of “body-self” and of our volition in its control.
Let’s return to our AI system. Through observation of pre-action intent states, actions of M, and their effect on X and Y states, the system develops concepts of body-self versus its environment, of its own volition over its body, and of its ability to influence the environment through body action.
At the same time, it has been observing Z states, modeling cause-effect relationships between Z states and others. These have a peculiar nature. Firstly, Z states have no direct power over the external environment— ie: Z states do not directly partake in cause-effect relationships within the environment. However, occasionally Z states have indirect power over the environment via the body-self — eg: a Z state may cause a particular M action, and that action causes an effect on the environment. Secondly, Z states have close correlations with X and Y states related to the “body-self” (eg: Z states that cause M actions leading to X and Y states, and X and Y states that cause Z states), but almost no correlations with X and Y states related to the environment (ie: few, weak, or only indirect cause-effect relationships between aspects of the environment and Z states). Thirdly, the model of the body-self has a clear boundary, and the source of its X and Y senses can be given clear locations on that boundary surface. This is possible because, for example, there is a consistent and predictable X and Y response associated with touching different parts of the body or eliciting different M actions.
In contrast, the Z sense has no clear location. No location on the body can be touched to predictably elicit a particular Z state. No M action can be produced with a consistent associated Z state. Sometimes Z state does correlate more strongly with X and Y states (eg: when attending to X in such a way that the cognitive system just produces an almost unchanged representation of X as its output, which is fed back in via Z), but even then its form is different to that of the X or Y that it correlates with. But at other times Z bears no relation whatsoever with X or Y (eg: when thinking). Lastly, whereas X and Y are associated with latent models both of the body-self and of the external environment, Z is only associated with a latent-model that is independent of the external environment (when considering only direct relationships).
Consequently, the system would conclude that a) Z informs about a latent state that is part of the self, b) the latent state correlates to no physical location on the body, and c) the latent state is the leading cause of M actions.
Eventually, the system would summarize that as a concept of “cognitive-self”, also known as “mind”.
Consequences for Understanding Intelligence
There are many interesting consequences to draw from the description of deliberation and auto-meta-management. I’ll discuss some consequences for our understanding of human intelligence first.
Evolution
There is a likely evolutionary path suggested by the interaction between deliberation, meta-management, and modeling. It is theorized that modeling capabilities first evolved in mammals to simulate possible futures (Bennet, 2023). The earliest mammals were very small and had many predators, and this simulation ability enabled them to sneak past their predators by considering the likelihood of whether the predators would notice them.
It’s possible that modeling and model-based inference evolved initially as part of non-deliberative simulation, or with such short deliberations that they could be managed through simpler means. The brain would simply do the best prediction it can on a single pass-through. With that modeling capability in place, and the evolutionary pressure to achieve more accurate predictions through longer deliberations, the modeling capability could have turned towards the internal cognitive behaviors.
Abstract thought
Deliberation almost certainly evolved for the purpose of better considering things about the environment for the better survival of the animal. With it evolved meta-management for the purpose of keeping that deliberation on task — ensuring that it produces useful results w.r.t. the immediate concern. Put in other words, meta-management exists to ensure that there is always a close relationship between the cognitive trajectory and the current physical state. But it does this through modeling of the cognitive state space, giving it an understanding of cognitive processes beyond how they relate to physical needs.
Consequently, it’s possible for cognitive processes to dissociate from immediate physical needs while still being useful to the organism. For example, the organism can identify some cognitive state goal (maybe to deduce how a prey got to its hiding place during a chase earlier in the day) and the cognitive modeling will help navigate cognitive state space in finding a path towards that goal. Later, that same cognitive modeling capability enables us to develop abstract concepts, to place them in cognitive space in relation to each other, and to navigate them in the production of thought.
Thought is complex and varied. The strategies we use to rationalize about things are learned. That learning would not be possible via model-free based systems alone because there would be nothing to provide the feedback to indicate whether one cognitive trajectory is better or worse than another. Self-modeling of one’s own cognitive space provides that feedback. It’s ironic that the system that evolved to keep cognitive processes tied to the physical needs also becomes the system that enables the processes to be completely removed from physical needs.
Thus, in attempting to understand consciousness, we have discovered something far more imminently useful — one of the key mechanisms behind humanity’s higher intelligence.
Consequences for Understanding Consciousness
Given the mechanisms described, in any system with similar levels of functional capability as human brains, the emergence of phenomenal consciousness is inevitable.
Consciousness has been studied for eons by philosophers, and more recently by neuroscientists, mathematicians, physicists, and computer scientists. So many questions have been raised. The various discussions fill many books, and many more will be written.
I believe that the meta-management theory of consciousness provides many answers, but I cannot hope to address them all. Instead, I will point out a few conclusions that I’ve found to be of interest.
Limited access consciousness
We are only conscious of a small amount of what goes on within our brains. Many brain processes occur without any conscious awareness whatsoever. Others provide only minimal access. For example, when consciously deliberating over something, we are aware of step. However, we are usually not aware of how each step is carried out. For most people reading this sentence, when I ask you what 5 + 3 is, you will either gloss over the question and continue reading, or you will find that the answer is simply there immediately while you continue to read. For the former, stop for a second and think. How did you get the answer? Many people will find that they simply hold the question in mind for a second, let the mental space open, and the answer appears. For something simple like 5 + 3, the answer is probably available from a rote-learned response — you simply recalled the answer from memory, rather than working anything out. But how does that memory recall work? We have no conscious access to the processes of memory recall. Many studies have shown that we are conscious of a tiny fraction of all that happens within our brain.
This is a direct result of the meta-management feedback loop. We are only conscious of what that feedback loop provides access to. That feedback loop only provides access to the minimal set of information needed to support effective deliberation, as optimized by evolution.
For that very reason, our conscious access is limited in two specific ways:
- scope — we only have access to those processes that we need to monitor. Anything else will simply be omitted entirely.
- granularity — the feedback loop captures only a high-level summary of the activity of processes that it does capture.
I described earlier how the different sensory modalities have differing levels of acuity within conscious experience, and I suggested that perhaps this is tied to those processes that require model-based processing. It’s interesting to note that the feedback loop does include information about the processing of our own cognitive state (ie: awareness of awareness) — otherwise we would not be able to theorize about our own consciousness. This suggests perhaps that meta-processes are sufficiently complex in order to need meta-management. That is plausibly true, although it leads to questions of infinite regression. Another explanation is simply that the meta-management feedback loop captures any cognitive processing that involves the model-based system. Since reaction to cognitive state requires model-based predictions, the processing of that state gets re-captured by the feedback loop.
In other words, it would appear that deliberation, model-based inference, meta-management, and consciousness all go hand-in-hand. They are likely all part of the same system.
The intentionality of consciousness
In the philosophical world, intentionality refers to the target of a representation. The intention of a visual perception of an apple is the apple itself. The intention of an abstract thought is whatever abstract concept the thought is about. The purpose of the term is to distinguish between what does the representing, and what the representation is about. Most perceptions are fairly easy to understand. We have visual, auditory, and tactile sensory organs that result in a predicted latent state (the representation) of the external thing that we are perceiving (the intention).
Conscious perception is strange. It appears to “look through” to the intentions of other representations. For example, our conscious experience of perceiving an apple has the apple as its intention. The Higher-Order Thought (HOT) theory of consciousness derives directly from this observation. It suggests that the brain constructs first-order representations based on raw perceptions, and then constructs higher-order representations about those first-order representations. Many have asked why the brain should do such a thing. If it already has first-order representations about the thing being perceived, what’s the point of generating additional representations?
Here we have the answer. And, as usual for the path of scientific discovery, the answer is somewhat different from the assumptions implicit within the questions.
The meta-management feedback loop captures certain aspects of cognitive processing and makes them available as a sensory input, like any other. If the immediately prior cognitive processing was w.r.t. to some visual perception, then the cognitive sense will capture something about that perception. It won’t be exactly the same as the original visual perception — it will be dimensionally reduced and altered in some way by that original cognitive processing. As all raw sensory inputs go through a predictive process to infer their latent state, the cognitive state that is considered for subsequent processing will be further augmented with model-inferred information about the source of that cognitive state (the “cognitive-self”) and other associated connotations.
I would suggest that the latent state inferred from that cognitive sense is what we think of as HOTs. The difference between key published ideas about HOTs and what I thus predict can be summarized as follows:
- HOTs indeed capture higher-order representations about first-order state. That state could represent thoughts or perceptions, with no fundamental difference between them (unlike the assumed difference between HOTs and HOPs — higher-order perceptions).
- Some have suggested that HOTs duplicate or perhaps even share the same underlying first-order representations. This is not correct. HOTs are a higher-level summary of the first-order states. This discards much information from the full first-order perception or thought that we are not consciously aware of. We have been tricked. We think that our conscious experience of our perceptions is an accurate representation of those perceptions. It is not. It is merely an impoverished approximation.
- Some have questioned why the brain should expend precious resources on the production of HOTs. Others have discussed whether HOTs are always produced, or whether they are only produced somehow “on demand” for those times when they are needed. The answer is twofold. Firstly, HOTs contain little information compared to the original cognitive state that they represent, so the cost of generating them is minuscule compared to the total brain function. Secondly, they are most likely generated continuously. The cognitive sense is very likely managed in the same that our brain manages other perceptions. Our eyes and ears continuously generate signals but those signals become attenuated if they offer no information beyond background noise. Even when our eyes and ears produce meaningful signals, we can still ignore that information if we attend to something else. I expect the same applies to the cognitive sense.
The causal nature of consciousness
Many have asked about the causal effect that consciousness has on brain function. For example, the conscious experience of a thought presumably coincides with the thought itself. Does the conscious experience contribute to the result of that thought?
Some experiments have found that the conscious experience of a decision to act may be delayed by as much as 10s (Soon et al, 2008; Haynes, 2013). The usual experiment works by asking a subject to take any amount of time that they like to make a decision. They will be in an environment with an accurate timing display in front of them — for example, a fast moving clock hand that indicates fractions of the second. They are asked to take note of the position of that hand the moment that they make the decision. At the same time, the experimenter records brain activity and records known tell-tale neurological signs of a decision having been made. By comparing the reported location of the clock hand and the brain activity, the experimenter is able to conclude that the brain activity of the decision occurred before the moment identified by the subject. There are still problems with the experimental design, but it has been repeated multiple times. It suggests that consciousness does not contribute to the decision. So what’s the point?
From a different point of view, the suggestion that the biology of brain function might be explained in mechanistic terms feels somehow disconnected from our experience of consciousness. To some, consciousness feels like something from another dimension that observes the goings-on within the mechanistic brain but doesn’t operate on the same principles. If that were the case, does consciousness have any causal power over brain function at all? Could it be epiphenomenal, a phenomenon that correlates to brain activity, but doesn’t partake in it.
The ideas presented in this post offer a mechanistic basis for consciousness itself and explain where it has causal power and where it does not. I refer to consciousness as post-causal: meaning that it has causal power over the events that occur after the perception or thought that it is about.
Consider any perception, thought, or decision, and for the sake of simplicity imagine that it occurs instantaneously, ie: within a single “iteration” of deliberation. I shall talk about a decision, but the same goes for thoughts and perceptions. Immediately following the decision, the feedback loop captures the cognitive state containing the decision and makes it available as a cognitive sense. On the next “iteration” of deliberation, the cognitive sense containing information about that decision is available for further processing — but it is too late, the decision has already occurred. By the time that the decision is being consciously considered (ie: the cognitive sense of the decision is being further processed), unconscious processes may already have been initiated to begin to carry out the decided motor action.
This fits with neurological data. To a somewhat poor approximation, cognitive processing is a time-share system. For it to both make a decision and to be aware of that decision, it needs to do that sequentially. First, it makes the decision. Then it does whatever it needs to confirm that it has sufficiently completed the prior task, and begins to consider what to attend to next. At that moment, it discovers mention within short-term memory that it has been asked by an experimenter to report when it makes the decision, and that it needs to use some other perceptual reference to identify that time. Finally, it switches to consideration of the relevant perceptual information and deduces the outcome to be reported verbally. It’s no wonder that there’s a 100ms or so delay with all that has to be done.
But that’s not the end of the story. This does not mean that consciousness is acausal or epiphenomenal. The cognitive sense provides important feedback about what has happened in the immediate past (as recent as less than 100ms). This enables the cognitive processes to reconsider their actions. For example, to veto or halt the action that has already started, or to generate a report about being consciously aware of taking the decision.
Consciousness is post-causal. It has significant influence over the mental events that occur in the future. It’s what causes us to go from observing an apple on a table to pondering about our own awareness.
The functional purpose of consciousness
So, consciousness is causal. That means that it serves a function, right? Not so fast.
When we ask about the functional purpose of some feature or capability, we are asking about a couple of things. First, if that feature or capability did not exist, would we lose a functional capability at the macro level? Second, in the case of biological organisms, what evolutionary benefit did the feature or capability confer? You’ll notice that these are very similar questions but the different phrasing helps to get a grip on the meaning of the word “function”.
At the beginning of this post, I started by discussing deliberation and its function. I continued by talking about meta-management, and its function. Deliberation and meta-management have clear answers to both phrasings of the functional question. The cognitive sense serves a key function within meta-management. Most further processing of that cognitive sense is for the purpose of meta-management during deliberation.
Consciousness is not the meta-management process, per se. It is the result of the cognitive sense and of the ability to further process it. It is possibly best described as the ongoing perception and memory of cognitive state while that is being used for meta-management. Thus, consciousness is an emergent property of that process, but it is not functionally required. If meta-management were to somehow occur without the emergence of consciousness, we would be missing no functional capability. We would in fact function identically.
So, phenomenal consciousness has no functional purpose. There is no evolutionary advantage to the phenomenal “feels” of consciousness over and above the correct functioning of the meta-management processes from which it emerges.
This leads to an interesting question: is it possible for meta-management to occur without consciousness? With a different architecture, yes. With the architecture of auto-meta-management via cognitive state feedback loop, no. I believe that, while consciousness is not the meta-management process itself, in practice consciousness is inseparable from meta-management in brains that follow similar architectures to humans. Put in other words, given the mechanisms described for auto-meta-management, in any system with similar levels of functional capability as human brains, the emergence of phenomenal consciousness is inevitable.
Consciousness by degree
You heard me right. Consciousness is a mechanistic process. It results from a particular functional structure. It is multi-realizable. Any replication of that functional structure will produce consciousness within that structure, regardless of substrate. We’re talking alien consciousnesses in exotic biological materials, we’re talking artificial consciousnesses in silicon, we’re even talking planet sized consciousness in semaphore flag wielding citizens.
We can begin to define some minimal conditions for the emergence of consciousness within any system:
- Modeling. The system must be capable of observing and modeling arbitrary cause-effect relationships.
- Feedback loop. The system must contain a feedback loop that captures its computational state and makes it available as a sensory input. That sensory input must be exposed to the aforementioned modeling capabilities.
- Computational capacity. The system must have sufficient computational complexity within its modeling capability and within its computational processing capability to “understand” its cognitive sense. This is somewhat vague, but I will return to it shortly.
- Evolutionary need. From a biological point of view, there must be evolutionary pressure for the organism to evolve the required mechanisms. For example, it is unlikely for consciousness to arise in organisms that lack deliberation, as they do not need meta-management or the structure that goes with auto-meta-management.
The last point suggests that some animals have conscious experiences while others do not and that there is likely a clear structural difference in the brains. If our current understanding is correct, modeling and deliberation evolved in mammals. As deliberation requires meta-management, all mammals likely experience consciousness. Consciousness in other vertebrates and in invertebrates is a question of whether they separately evolved deliberation. For example, there is evidence that some spiders may have the capacity (also see Cross & Jackson, 2016).
What might conscious experience be like for these different species? Would it be as rich and meaningful as it is for us? Likely not. The informational content of any raw cognitive sensory input depends on the amount of scope and detail of information extracted by the meta-management feedback loop. The scope and detail of each newly generated cognitive state depend on the level of complexity and accuracy captured by the model and on the computational complexity of the cognitive processes. This is what I call the computational capacity of the system, and for any given organism it is optimized by its specific evolutionary needs.
For any organic and artificial system with consciousness, the richness and acuity of its conscious experience come in degrees, according to the computational capacity of the system.
Notice that this conclusion is consistent with attempts to calculate objective measures for consciousness. Integrated Information Theory, the most famous example, proposes that consciousness only occurs in certain systems with certain functional structures and that its degree varies by the amount of mutually causal information. The less well-known Information Closure Theory of consciousness (ICT) has a similar basis in informational content and draws the same conclusions about the importance of structure in determining the necessary conditions for consciousness and its degree.
What consciousness “feels like”
The question of why consciousness feels the way it does has led to so much debate that is it now well known as the Hard Problem of consciousness. The “feels like” nature of consciousness was first described by Thomas Nagal when considering how it may feel different to be a bat than what it feels like to be a human.
In practice, what we mean by the “feels” of consciousness is pretty vague. I think what researchers mean today is this:
- Presumably, all the mechanical and silicon machines that we have ever built as of today have no conscious experience whatsoever. If we scaled up one of those large enough, it could in theory emulate intelligent thought. Perhaps it could even behave completely like a human and could have access to its own internal computational processes. However, it would still just be a scaled-up version of an unconscious machine.
- Intuitively, there is a difference between the characteristics of our own conscious experience — the “feels” of consciousness — and whatever characteristics of informational content that any machine could have.
I beg to differ.
Firstly, while some will disagree with me, I will state assuredly that all conscious experience carries informational content. The “feels” of consciousness is information, no matter how vague. That information has a computational basis — something obtained or generated that information. Humans feel consciousness to be the way that it does because:
- All sensory input is processed w.r.t to models to infer their latent state, re-constructing a guess about what caused the raw sensory data. This adds information that was not included within the raw sensory data. Such additional information includes all sorts of things like our belief about where the sensory perception came from, and what it means to us.
- When queried, our memories attribute further information, fleshing out further and further the “meaning” associated with any given perception.
- Our emotional affect additionally adds information, painting our particular current emotional hue over the latent state inference that is made from the raw sensory data.
- All of that occurs against our exteroceptive senses. There’s good reason to believe that it occurs against our interoceptive senses just the same (Feldman et al, 2024). The same goes for our cognitive interoceptive sense.
Remember, our exteroceptive perceptions are not accurate renditions of the world outside. Our conscious access to our cognitive processes is not an accurate rendition of those cognitive processes. In all cases, our sensory perceptions are muddled guesses, colored by our own expectations. The “feels” of consciousness is simply whatever is inferred from that noise.
I challenge you to find flaws in this argument.
Furthermore, with the “feels” of consciousness explained via mechanistic account, what remains of phenomenal consciousness that still has no mechanistic explanation? This is another challenge to the reader and those in the community at large.
Closing Remarks
The idea is not new that consciousness is intimately tied to meta-management, or to meta-cognition more generally. I don’t consider myself to be the inventor of the first “meta-management theory of consciousness”. But if you try googling “meta-management theory of consciousness” today you won’t find any other exact matches. It has been mentioned in passing by other researchers, but not thoroughly investigated.
I believe what I have presented here is the clearest explanation to date of how meta-management relates to consciousness. It provides a clear evolutionary context, proposes a mechanism for meta-management that fits with the phenomenology of consciousness, and provides clear explanations for the mechanisms underlying that phenomenology. Furthermore, it offers insight into human intelligence more generally.
So I’d like to claim that it is not just the clearest explanation of how meta-management relates to consciousness, but that it is the clearest explanation of consciousness overall. But only the community at large can decide that.
Do reach out if you want to discuss further any of the ideas presented here. You can comment on this blog post if you have a medium account. Other ways to contact me are available on my website:
The website also contains a significantly more detailed analysis of the ideas presented here:
References
Bennet, M. (2023). A Brief History of Intelligence: Evolution, AI, and the Five Breakthroughs That Made Our Brains. Mariner Books.
Beaudoin, L. (1994). Goal processing in autonomous agents [PhD thesis]. The University of Birmingham. Full text: https://citeseerx.ist.psu.edu/document?doi=382135c4379c08253810ef8f5823c469af6b69df
Cross, F. R., & Jackson, R. R. (2016). The execution of planned detours by spider-eating predators. Journal of the experimental analysis of behavior, 105(1), 194–210. https://doi.org/10.1002/jeab.189
Cushman, F., & Morris, A. (2015). Habitual control of goal selection in humans. Proceedings of the National Academy of Sciences of the United States of America, 112(45), 13817–13822. https://doi.org/10.1073/pnas.1506367112
Dolan, R. J., and Dayan, P. (2013). Goals and habits in the brain. Neuron 80, 312–325. https://doi.org/10.1016/j.neuron.2013.09.007
Douskos, C. (2018). Deliberation and Automaticity in Habitual Acts. ETHICS IN PROGRESS, 9(1), 25–43. https://doi.org/10.14746/eip.2018.1.2
Feldman, M. J., Bliss-Moreau, E., & Lindquist, K. A. (2024). Trends in Cognitive Sciences. https://doi.org/10.1016/j.tics.2024.01.009
Haynes, J-D., 2013, “Beyond Libet: Long-Term Prediction of Free Choices from Neuroimaging Signals”, in A. Clark, J. Kiverstein and T. Vierkant (eds.), Decomposing the Will, Oxford: Oxford University Press (online edn, Oxford Academic, 26 Sept. 2013). https://doi.org/10.1093/acprof:oso/9780199746996.003.0003
Soon, C., Brass, M., Heinze, H-J., and Haynes, J-D. (2008). Unconscious determinants of free decisions in the human brain. Nature Neuroscience, 11, 543–545. https://doi.org/10.1038/nn.2112
