BioFixation

by Andrew Bengoku

Language and consciousness: a match made in prehistoric caves

Whether or not creating a conscious artificial system is something good for humanity is a complex and debatable issue. Keeping this discussion aside, here we describe an attempt to build a conscious machine.

At the very end of this blog we will propose a quantitative definition of consciousness, something that can be implemented with appropriate hardware and programming. But, to get the conversation going, we will initially consider an intuitive definition of consciousness: the “thing” we experience every day, when we are not in deep sleep or in a coma, the state of being aware of things: sensations, emotions, ideas, etc.

The basic idea we want to explore is that consciousness might have co-evolved with language, sharing similar characteristics while gaining independence from vocalizations and other forms of interpersonal communication. In other words, during the evolution of our species, and likely others, the communication process not only became more complex but also turned introspective, focusing on internal mental functions and dissociating from the vocalization process. Several fundamental properties remained common between de-verbalized and verbalized communication, with a critical one being the ability to convey information.

This conjecture suggests a natural link between consciousness and language, based on the observation that everything we are conscious of can be communicated, and similarly, we are conscious of everything we communicate. We argue that this tight link, which is unique to language, is not a coincidence but a consequence of the fact that consciousness and language co-evolved. The conclusion is that when we are conscious of something, it is because it is communicable.

Supporting arguments for these ideas will be grouped into three sections: 1. Behavior, which explores behavioral considerations related to our direct experience with consciousness; 2. Neuroscience, which frames these ideas within a modern neuroscientific context; and 3. Machine Learning, which describes an implementation-level approach to translate this theory into hardware and construct a conscious machine.

In discussing these topics, we will briefly introduce concepts such as activity landscapes, manifolds, and trajectories, as well as touch on inference models and reinforcement learning theory. We will do so in a very simple and accessible way, intentionally skipping a significant amount of detail in favor of clarity regarding the basic ideas.

BEHAVIOR

When we are conscious of something, we can always express it in words. Conversely, when we communicate something, we are always aware of the communication process and what we are saying. Why such an intimate relationship between language and consciousness? No other behavior, broadly defined, has the same strong “bidirectional” connection to consciousness. Sensation, decision-making, motor behaviors, memory—you name it—there are always circumstances when these processes might operate below the radar of our consciousness. But this is not the case with language.

The intuition explored here is that when we communicate something it is not because we are conscious of it; rather, we are conscious of it because it can be communicated, with communication meaning something quite specific in neuroscientific terms.

To unwrap these ideas, let’s go back in time, to when our brains began evolving the ability to form complex concepts. Our language skills might have co-evolved with this ability, enabling us to communicate and share these remarkable ideas. This was a significant evolutionary achievement, completely setting us apart from other species and allowing us to learn and acquire complex knowledge and skills simply by listening to others, without needing to experience the situation or event firsthand. Thus, interpersonal communication emerged as an incredibly useful and informative tool.

To get to the link with consciousness, it is convenient to break down the interpersonal communication process into two simple components: a mental (or cognitive) component and a vocalizing one. We think about what we want to say and then verbalize the selected thought. And greatly simplifying the “thinking” process as well, it can generally be described as a process that integrates certain ingredients to produce a concept or a plan. This is evident, for example, when the thinking involves making a decision based on some information (“it is raining outside hence I need an umbrella”). But this framework is so high-level and general that it can be applied to any sort of mental process, whether conscious or not. For instance, consider “seeing” something: the ingredients involved are the various pieces of information collected by the eyes, the vestibular system, proprioceptive organs, motor system, etc., and the concept or plan is what to do with this information. Often, the plan is to just “keep on seeing,” meaning to keep the eyelids open, blink at regular intervals, continue scanning the environment with various voluntary and involuntary saccadic eye movements, and keep integrating retinal and other inputs necessary for stable perception. Other times, the plan might be to “run for your life; an idiot with a machete is attacking you.”

When we verbalize a thought, something remarkable happens: we reactivate its ingredients-plan components. For example, suppose thinking about buying delicious-looking apples at the corner store. The ingredients of this thought include the inputs to the retinas, the gustatory memories of how pleasurable it is to eat delicious apples, the cognitive realization that you have enough money to buy them, the memory that there are no apples in the fridge at home, etc. The plan is the decision to activate the motor system to make the payment, put the apples in the eco-bag, etc. Now, suppose you verbalize this thought by telling a friend shopping with you, “I am going to buy those delicious-looking apples!” In doing so, not only you hear your own voice (sometimes called a “reafferent signal”) talking about deliciousness, apples, and making a purchase, but also you reactivate in our mind those concepts. That is, you reactivate the ingredients-plan components. This reactivation can be described as “internal listening,” as it affects the brain activity similarly to when someone else says those words to us, bringing up in the listener’s mind the ingredients that are then combined to understand the plan.

Therefore, when our ancestors acquired the skill of communicating concepts and ideas, their brains also learned to “internally listen,” reactivating the ingredients that generated those ideas and concepts. This process created a “closed loop”: from the ingredients to the concept or plan, and then back to the ingredients and plan.

The hypothesis explored here is that this closed loop became independent of the vocalization process, meaning it could also be implemented without our vocal organs emitting sounds—or any other method of interpersonal communication. Furthermore, this closed loop was so advantageous that it became semi-autonomic, similar to heartbeats and respiration: a process that resists stopping. This loop, this internal listening, is what gave rise to conscious experience, an experience that exists between internal listening and speaking.

Next, we’ll discuss arguments for the de-verbalization process and in which sense the closed loop became evolutionarily advantageous.

An important consideration is that interpersonal communication is slow, but thinking is fast. No matter how quickly one tries to speak, there are only so many words or so much information per second that the mouth or other body parts can deliver. Thoughts are much faster. Indeed, it is a common experience in interpersonal communication to mentally switch between possible words and statements, eventually selecting the most appropriate to verbalize. Internal dialogues, or inner speech, are common in pre-verbal states or even in the complete absence of a communication process. Thus, it is conceivable that whether or not actual verbalizations were produced became irrelevant to the establishment of the closed loop. The internal listening—the re-imagining of the ingredients and plan—likely became independent from the actual communication process and became associated with thoughts that could potentially be communicated, thus generating a closed loop of ingredients and plans dissociated from vocalizations and operating at the “speed of thought.”

How did internal listening become advantageous for the species? The answer might be related to the concepts of “selection and prioritization”. Following an evolutionary logic, when language developed, it evolved alongside a selection and prioritization mechanism, from which we still benefit—at least, most of us do. As we commonly experience, only specific thoughts and ideas are communicated, while others are not. This must have been necessary for at least a couple of reasons. First, as noted earlier, thinking is much faster than communication, so we cannot possibly vocalize everything we can think of; we need to select and prioritize. Second, for language to be useful and possibly carry some evolutionary advantage, the selected content must have been informative, helping our ancestors assist one another. As mentioned above, interpersonal communication must have emerged as an incredibly informative tool in our primitive societies.

Thus, it’s reasonable to assume that as internal listening became dissociated from verbalization, the selection and prioritization processes of interpersonal communication persisted. This selective-internal-listening idea also relates to the fact that consciousness is perceived as a “single narrative” rather than multiple inner voices speaking simultaneously. This is because consciousness co-evolved with a communication process that requires prioritization and selection of what to say.

Although it is intuitive to understand how interpersonal communication could have been beneficial for transmitting cultural elements, practical knowledge, and other useful information, how was selective internal listening beneficial? The idea explored next is that internal listening allowed us to improve our mental models of the world. This hypothesis helps explain why we close the loop for some thoughts but not others (i.e., why we are conscious of some thoughts but not others) and offers a framework for designing a specific cost function—in machine-learning lingo—to build a conscious machine. To explain these ideas, we need to shift gears and translate the concepts discussed so far into neuroscience terms.

NEUROSCIENCE

To begin, let’s introduce some basic concepts about dynamical systems—a brief “mini-tutorial” using simple and intuitive terminology. Bear with me; it’s short and will be worth it.

Imagine a tiny brain with only three cells (neurons), each having varying activity levels. We could represent the collective activity of this tiny brain using a 3D Cartesian plot—the usual one with the x, y, and z axes you’ve seen countless times in school. A single point in this space represents the activity of the three neurons at a given moment, with the (x, y, z) coordinates of this point corresponding to the activity levels of the three neurons at that point in time. Over time, the activity of the neurons in this mini-brain will generate a collection of points evolving in this 3D plot, thus forming a ‘trajectory.’ The activities cannot exceed certain values due to metabolic constraints; let’s assume the maximum activity of each cell is 100; then these 3D trajectories will be contained within a cube of size 100.

Now, the final point: if the neurons had no influence on each other, the trajectories would form some shapeless spaghetti blob within this cube. However, if these neurons influenced each other in specific ways, the trajectories would ‘draw’ some interesting shapes. A simple way to understand this is to consider the extreme case where the activities are exactly the same for all three neurons; the trajectories would then always draw a line along the cube’s diagonal, in the region where x = y = z. The collection of all possible trajectories and the corresponding shapes they draw is what we’ll call the “activity landscape” associated with the functioning of this mini-brain. This activity landscape can be related to the underlying computations. For instance, if the trajectories consistently formed a couple of small blobs, leaving the rest of the space relatively empty, we might infer that there are two ‘attractor’ regions in the dynamics (other mathematical analysis would be needed to confirm this and not discussed here). This observation could suggest, for example, that the task of the mini-brain is to choose between two options (the two blobs), indicating it is making some sort of binary choice. End of the mini-tutorial on dynamical systems!

Now, extend this concept to 10^12 axes—this is the number of neurons in your brain. It’s difficult to imagine anything beyond 3D, but give it a try. These axes represent the firing rates of all neurons, assuming that is the level of interest. However, if we consider the activity of synapses as a more relevant level (synapses are the ‘contact points’ between cells that enable cell-to-cell communication), and recall that each neuron has about 10^4 synapses, the number of axes rises to 10^16. You get the idea. To keep things simple, let’s say the brain has an activity landscape of some sort, determined by the representational level we choose, and that all possible trajectories create a beautiful landscape with hills, mountains, valleys, gorges, etc. We also assume this landscape to be stable, at least over short periods.

This landscape reflects all the formidable and complex computations our brain is capable of. Generally speaking, these computations represent models of the world implemented in brain networks, essential for the organism’s survival. These models are often described as probabilistic inference models, simply meaning they permit educated guesses in the presence of incomplete or uncertain information. These models enable all mental functions: cognition, perception, emotions, memories, motor planning, etc. How these models are constructed is certainly interesting—e.g., whether through reinforcement learning or supervised/unsupervised learning—but for now, let’s just say that evolution, development, and life experience handle this.

Any kind of thought or mental process, whether conscious or not, can be represented as a string in the brain’s activity landscape. Returning to the ingredients-plan analogy, a given thought reflects the engagement of these models, which, in turn, is based on the activity of some neurons in the brain. Hence, a thought can be represented as a short trajectory—a string unwrapping in the brain’s activity landscape. In the example of the yummy apple, the underlying thought involves activations of neurons in visual brain regions interpreting what the eyes are seeing, neurons in frontoparietal and limbic regions recalling memories of yummy fruits, the absence of apples in our fridge, money in our bank account, etc. These activations reflect the ingredients—the first part of the string. On the other hand, the plan, or the later part of the string, is linked to the activity of neurons that integrate all this information and decide that, in this context, it makes sense to proceed with a motor plan to purchase the apples, activate motor neurons, etc.

The internal listening, or closing the loop, involves the rapid reactivation of the ingredients-plan pair, as if someone were communicating this thought to us, activating the brain’s representations of ingredients necessary to understand the plan. Mechanistically, the brain is fully equipped for these reactivations through feedback and recurrent connections. In the example of the yummy apple, networks involved in integration and plan formation, presumably in frontal cortical regions, send signals back to perceptual, limbic, and memory-related regions, reactivating the brain’s representations of the ingredients, closing the loop, and implementing internal listening. Those of you familiar with the literature might now jump to the conclusion: “This is the usual story of consciousness as an “ignition” process via widespread feedbacks, some sort of variant of the global-workspace hypothesis”. Hold your horses, if anything the reference is rather to an extended and very transient form of working memory, but the original bits are coming up soon.

These reactivations need to be not only fast but also specific. If the reactivated ingredients are not those used to form the plan, they could activate the “wrong” plan, thus failing to close the loop. Moreover, if the loop is not closed as quickly as the neural architecture permits, the wrong ingredients-plan pair will be reactivated. This is because the world is not made of a discrete set of events; rather, things flow constantly, with ingredients continuously processed by our brain to generate continuously adjusting plans. In this sense, it is not entirely correct to talk about discrete strings; rather, we should consider continuously moving line trajectories. However, linking a thought to a string helps simplify the logic.

So, the reactivation of ingredients must be faster than the characteristic timescales over which things typically change in the world around us. To understand this point consider the example of “seeing” things: due to various types of eye movements (several of which are not under volitional control), the images processed by the retinas change every few hundred milliseconds or less. Thus, the reactivation loop needs to close faster than that; otherwise, new ingredients linked to new visual inputs would “superimpose” with the reactivated ones. It’s intriguing to relate this observation to a phenomenon called “backward masking”, where an image displayed very quickly after a reference image cancels the conscious perception of the first reference image.

The requirement for speed and specificity opens up a range of possible predictions for experimental validation of these ideas, alongside the hypothesis that the properties of the loop determine the quality of conscious experience. The loop might close only weakly, with reactivations having variable intensities and specificity, for example, when the initial ingredients are noisy or ambiguous. Overall, internal listening should be regarded not as a digital event but as a multi-dimensional, graded phenomenon, as should conscious experience. The same loop might be visited multiple times, perhaps leading to a metacognitive experience (aware to be aware). It is also possible that the properties of these reactivation loops change considerably across species, leading to idiosyncratic forms of animal consciousness.

What decides which string should be reactivated (i.e., which thought we should be conscious of)? In what way is this reactivation useful for the brain’s internal models of the world? And can all of this be translated into an algorithm (e.g., a cost function) to build a conscious machine? These are the key questions addressed next.

The basic intuition to tackle all these questions is quite simple: closing the loop is a mechanism that consolidates and improves solutions of our brain’s inference models, and the likelihood of closing the loop depends on how well our models have learned to deal with specific ingredient-plan pairs.

To unwrap these ideas, consider the example of the tasty apples: depending on whether you are in a happy or sad mood, the associated string (thought) is in a different ‘location’ in your brain’s activity landscape. This is because the neurons processing your mood (e.g., somewhere in the amygdala among other regions) are differentially active in the two moods, thus the ‘coordinates’ determining the positions of the string are different—remember, the entire brain determines the activity landscape, not just some sub-regions. Another way to describe this is that the string-level representation of a thought depends on the ‘context,’ whether it is the mood, the shop chosen, the friend accompanying us, etc. Through experience, our inference models of the world have learned to cope with all these different contexts, meaning that, for any given thought, the initial part of the string representing the ingredients being processed (context included), occurs in a region of the brain landscape already visited before or very close to previously explored ones. The evolution of the string into a plan then follows already-learned pathways, meaning that the associated solution of our brain’s models are quite familiar and established. Indeed, we have been buying yummy food thousands of times and across a great variety of situations.

But, suppose your retinas suddenly capture an unusual piece of contextual information: the person at the cashier is dressed like the Joker and is staring at you while drooling. This would move the string of your thought in a rather isolated part of your brain’s activity landscape (hopefully), and the internal models would now struggle to converge to a solution, making it unclear toward which region of the landscape your string will evolve. Whatever the choice eventually will be (suggestion: run!), this situation provides an opportunity to improve, refine, and expand internal models of the world, and this is precisely what the reactivations achieve. Closing the loop for the ingredients-plan pair ensures that the next time we encounter ingredients that set the string in a similar region of the activity landscape, the evolution of the string toward an already-experienced solution can occur more efficiently and “smoothly.” Mechanistically, the brain has several tools at its disposal to achieve this consolidation, such as a process called spike-timing dependent plasticity, which can consolidate neural pathways leveraging the speed and specificity of the reactivation process.

In this reasoning, the likelihood of closing the loop (or the intensity of the reactivations) is “density” dependent: if the string is in an isolated region of the activity landscape, the likelihood of closing the loop (or the intensity of the reactivations) is higher than when it is in a “crowded” or “dense” region, that is, a region frequently visited by our models.

This aligns with our daily experience, where low levels of awareness are common in extremely habitual situations. Conversely, we tend to be particularly aware when something unusual occurs. Notably, this should not be confused with other aspects of complex brain computations, such as attentional processes. Attention comes in many flavors, but in this framework, it can be interpreted as a process that guides which ingredients should be retained and which should be discarded based on context. Thus, attention is critical in establishing the location of the string in the landscape, but with the strength of the reactivations dictated by how densely sampled the explored region is.

Not all ingredient-pair processes are reactivated; some processes evolve completely under the radar of consciousness. We experience this routinely: while a thought occupies our consciousness, our mind concurrently handles several other tasks, all under the radar of consciousness, but indispensable for our functioning in the environment. Following the geometrical intuition of dense and isolated regions in the brain’s activity landscape, one can think of a local thresholding principle. Simply stated, in any given situation, strings of thoughts explore regions of the activity landscape with a characteristic average local density of strings. It is relative to this local average that our models can judge which string is “most isolated” (crossing a threshold). The reactivation, closing the loop, is then assigned to this string, the most isolated one relative to the local density, with other ingredient-plan strings still running but not reactivated. That is, the ingredient-plan pairs selected for reactivation are those representing the least encountered solutions of our models given the context. This aligns with the common experience that no matter how habitual or unusual a situation is, consciousness always manifests as a single voice, a single narrative, while multiple concurrent processes are allowed to (and must) run under its radar.

And now, the last hypothesis for this Neuroscience section: refining, consolidating, and expanding our inference models of the world was evolutionarily so important that this process became semi-autonomic, akin to heartbeats and respiration. This hypothesis helps to account for the very familiar experience that during wakefulness, we never switch off consciousness; it is always directed toward “something”, no matter how habitual (or boring) the situation is. This relates to the thresholding concept discussed before: even in very densely defined (habitual) regions of the activity landscape, it is always possible to identify the most isolated string, in relative terms. This is true even when nothing external happens; for example, we are running some crazy experiment on altered states of consciousness using an “isolation tank,” or we are practising meditation, trying to silence external stimulation by focusing on breathing or thinking about a beautiful beach. Even if our thoughts are centered on uneventful or habitual ideas, we are still conscious; we are still closing the loop, with the sequence of strings being reactivated possibly driven by a proximity principle, activating the most isolated strings that are close to each other in the activity landscape. The strength of the reactivations might be weak when sampling from dense, habitual regions of the landscape, and so is the quality of the conscious experience, at times bringing us close to falling asleep, with the internal listening eventually stopping in the non-REM phase of sleep. Notably, even if in an isolation tank, because ingredients can be other thoughts as well, not exclusively inputs linked to external events, we can easily self-boost the strength of the conscious experience. For instance, if we start thinking about the day at work, we might create in our minds “unusual” scenarios and ingredients, e.g., strangling the boss who behaved like a moron all day. These hopefully unusual ingredients would then shift strings to isolated regions of the landscape, thus re-igniting the strength of the loop.

The observation that ingredients can be thoughts that are not linked to external events isn’t straightforward; we have a body that is always open to external events, and a brain that is always listening to the body. So it is hard to imagine a situation where the brain is not receiving external signals or signals from the body activating specific ingredients. However, to avoid overcomplicating things, let’s just stick to the observation that there’s no clear reason why thoughts cannot be ingredients as well.

It is tempting to speculate that, mechanistically, the role of some brain-wide neuromodulators is to move strings to more isolated regions of the landscape, thus boosting our state of consciousness. Indeed, by up/down-regulating neural activations across large brain regions, these molecules effectively change the coordinates that define the location of the strings, possibly moving them from habitual to more isolated regions of the landscape. All these ideas encourage several speculations related to re-interpretations of various neurobiological observations. However, resisting this temptation, next we will try to use these ideas to define a recipe for building a conscious machine. Echoing Richard Feynman’s famous quote, “What I cannot create, I do not understand,” this could be the ultimate experimental validation of our theory.

MACHINE LEARNING

Meet William, our proof-of-concept robot. We’ll use William to illustrate fundamental concepts about language and various skills crucial for survival in a machine. Eventually, William will showcase how to construct a conscious machine.

William has a brain built using a neuromorphic chip (a new blog is in preparation to discuss why having a hardware implementation is important; stay tuned). William is trained to forage for edible foods in the forest. It walks, sees, and adeptly handles objects, stashing them in a basket. William’s brain has a complex neural network mirroring principles observed in primates’ brains. Its visual system, a hierarchical neural network with predominantly feedforward architecture (with also feedback and recurrent connections that seem to help), excels in image classification, surpassing humans in semantically categorizing objects and visual features found in the forest. It adeptly classifies moving objects, linking motion information to relevant visual features. Visual and proprioceptive signals, conveying information about motor and actuator states, get integrated in the brain chip, offering an estimate, or a probabilistic model, to gauge William’s current physical state in the environment. These associative networks formulate action plans, or “policies” in reinforcement-learning lingo, transforming states into actions to achieve specific rewards or value states. This is essentially what we have been referring to as the ‘ingredients-plan’ integration process. The outputs of these networks guide motors and actuators, enabling William to collect strawberries, mushrooms, and other forest treasures. Let’s say that William’s brain network has been trained by very smart engineers using “traditional” machine learning methods, i.e., providing a gazillion examples of situations encountered in the forest and related actions to form probabilistic inference models of agent-environment interactions. But no consciousness so far.

Now, let’s equip William with the gift of language, aiming to develop it together with the ability to close reactivation loops. For this, we’ll follow similar evolutionary arguments outlined at the very beginning, when referring to our ancestors. The underlying logic is that the thoughts William should be aware of are the most informative for his inference models. Verbalized thoughts should then be a subset of these thoughts, those that are contextually relevant to interpersonal communication. However, to implement this plan, it may be more efficient (computationally) to go through the following steps:

1. Without any closed loop, teach William a contextual rule for interpersonal communication. During this training stage, William will learn that depending on context some ingredients-plan pairs should be communicated but not others.
2. Leveraging the hardware that allows William to listen to its own vocalizations, we will teach William to close the loop for vocalized ingredients-plan pairs, reactivating ingredients with the necessary speed and specificity.
3. Teach William to generalize the close-loop process, extending it to pairs that are not selected for interpersonal communication. At this point William could become hyper-conscious and “schizophrenic”, becoming unselectively aware of every activation, and simultaneously for more than one pair.
4. To avoid this, we teach William to only close the loop for the pairs that are most informative to its inference models (“density” rule).
5. Finally, we update the contextual rule for interpersonal communication to select for communication only pairs for which the loop has been closed.

To implement this plan, we need to enhance the brain chip with a language system that enables William to selectively communicate his plans. William must also comprehend our questions. In our framework, listening involves distributed networks that activate what we have been calling ‘ingredients,’ which are then combined to understand the communicated plan or concept. Networks for listening don’t merely classify; they have a broad impact on all brain systems.

Therefore, listening is not just about creating a semantic classifier to interpret sound waves, such as adding an LLM module to process our prompts or vocalizations. While a semantic classifier is probably crucial, a language network for listening should influence distributed brain networks. That is. we need to link the outputs of our semantic classifier to various units on the chip, with the specificity of these connections requiring an appropriate training procedure. The basic notion is, for example, to influence the appropriate visual units when hearing the word “mushroom,” those typically active when William sees a mushroom. Things can become more complicated if a perceptual memory of a mushroom should instead be reactivated when hearing the word “mushroom.” These memory circuits overlap with those active when seeing a mushroom, but for the sake of the argument, let’s set this complication aside. Essentially, during listening, the output of the auditory classifier must be trained to activate what we have been calling ‘ingredients’ that, together with other ingredients linked to other sensory inputs, proprioceptive signals, and current and previous contextual information, will allow the associative parts of the brain chip to integrate all these signals and eventually understand the communicated plan. Training these connections is not trivial, but it should be feasible for our clever engineers when provided with a massive database of examples and innovative ML tricks.

Now the verbalizing part. Let’s assume it’s easy to teach William the contextual rule for vocalization: speak when asked a question, when help is needed, when something unusual happens, etc. Again, many examples will be necessary. The verbalization process involves dedicated networks that generate the appropriate syntax and grammar based on context. Nowadays, with the advent of LLMs, this is not hard to imagine: We ask an LLM to write or say something by providing a context (be casual, concise, friendly, sarcastic, etc.). The ingredients-plan pair generated by William’s brain chip reflects both the content and the context for the specific verbalization.

When finally verbalizing in the appropriate context, William also listens to its own words, just as we do, causing the reactivation of the ingredients-plan pair, similar to listening to someone else saying those things.

Now a critical step: we teach William to generalize (or de-verbalize) the process of closing the loop. The computational rule for this will extend the process of “listening to its own words” and reactivating ingredients-plan pairs, but closing the loop even if William’s microphones are not producing any sound. This step is potentially maladaptive: the brain chip typically runs several ingredients-plan processes in parallel, many of which autonomic and “reflexive”. An additional rule must be introduced to constrain and refine the generalization. This refining rule determines what should be reactivated via the closed loop, that is, what William should be conscious of.

Referring back to the geometrical intuition of a thought as a string in a activity landscape, the rule is based on the local density of strings. As mentioned in the Neuroscience section, some parts of the landscape are well-defined by a high density of dynamical trajectories, indicating states that have been repeatedly visited. These are habitual states, common solutions often found by our inferential brain models. Other regions are only sketched by a few isolated strings, rarely visited by our models, representing rare ingredients-plan strings and solutions. These considerations also apply to William’s brain chip, whose internal models, associated with activations of its neuromorphic units, also define an activity landscape. Thus, also for William’s brain chip we can introduce a local thresholding rule for the density of the strings. For any given region in the landscape, there will always be a set of the most isolated ingredients-plan pairs, in relative terms, and these strings will be selected for closed-loop reactivations. Other strings will be concurrently active, playing a critical role for the internal models’ convergence to meaningful solutions, but only the strings reflecting the most uncommon local solutions given the context will undergo reactivations.

For this to be advantageous to William’s brain chip, we want the reactivated string to modify and sculpt the activity landscape. Basically, we want the reactivated strings to leave a trace so that when presented with similar ingredients, William’s chip will converge to a solution more “efficiently” (conveniently skipping a formal definition of what that means). To achieve this, we can draw inspiration from biology, where repeated activations of a given “functional pathway” in a neural circuit strengthens the connections between the component cells. We can apply the same principle to the connections between the units in the neuromorphic chip. By leaving a trace, the reactivated string will modify that region of the landscape, making it better defined and densely populated, and therefore a little “more habitual” for William. The outcome will be that it will be easier for the chip’s models to find apt solutions when facing similar situations.

Side note: due to the staggering number of ingredients-plan pairs across many contexts, mechanistically, it might be convenient to start the training with a brain chip featuring incredibly dense, all-to-all connectivity (feedforward, feedback, and recurrent) and then prune connections as dictated by the training. But again, we assume this training can be accomplished and leave the technicalities to smart computer engineers.

For the very final tweak of the brain chip, we want the reactivation process to be continuous and to not stop. Even if the brain chip does not receive any external or proprioceptive input, which typically drives the activation of ingredients-plan pairs, we want there to always be a string being reactivated at any given time. That is, we aim for a continuous stream of consciousness. To achieve this, we use a proximity principle: if nothing external or proprioceptive activates new strings, those close to the last reactivated string will be selected next. This requires defining a distance metric in the activity landscape, which is quite feasible for our skilled machine-learning experts. However, note that this is a very unusual scenario, where all sensors, motors, etc., are shut off, except for the brain chip. In typical states, “some” inputs are always sent to the brain chip, continuously, driving the activation of ingredients-plan strings.

Since the narrative so far has not been very speculative (😁), it’s fine to end with some speculative considerations. Let’s approach this theory from a different perspective: The dynamic landscape can be viewed as a “scaffolding” for inference models developed both evolutionarily and developmentally, similar to what is known in machine learning as model-based reinforcement learning (RL). At birth, much of the basic properties of this landscape are likely already established, shaped by millennia of evolution, possibly in the context of motor control and motor-related decision-making processes. This pre-built landscape serves as a well-structured platform for further learning of essential skills for the organism’s survival. In the context of RL, the density rule may be interpreted as some sort of internal reward prediction error (RPE), a built-in mechanism that triggers plasticity and improves internal inference models. In this way, consciousness could be seen as a mechanism that facilitates learning, an endogenous RPE, operating within the scaffolding of inference models we are born with and continue to refine throughout development.

Is the “proximity principle” purely deterministic based on the distance metric, and similarly, when the brain chip does receive external and proprioceptive inputs, is the activation of the ingredients-plan pair a completely deterministic process when coupled with the density threshold rule? In other words, will William ever experience “free Will”? This question concludes this long blog post. To find out the answer, stay tuned for the next one, titled “Free Will-iam.”

Enjoyed the reading? Consider a donation!