Over the past year I have become fascinated with the human mind. Originally piqued by Karl Friston’s free energy principle (FEP), my interest intensified as I began to learn the mechanisms of the mind, and how neural architecture impacts my perception and cognition. I believe that understanding how we think on a mechanistic scale is necessary, if not sufficient, for unlocking the full potential consciousness affords us. Thus far I have only begun my journey into my own mind, but I would like to propose a framework unifying several relatively nascent ontologies.
I will start with an abstracted, anthropomorphic representation of the mind, that while leaky provides an easily comprehensible framework to build on. Following this, I will outline two mechanistic models of the brain. Finally I will attempt to tie these ontologies together intuitively by showing how they build on each other to create a cohesive cognitive model.
An anthropomorphic representation of the mind
To build up this explanation of the mind, two disjoint models need to be somewhat understood.
Stanislas Dehaene proposed in 2011 a theory of cognition and consciousness known as Global Neuronal Workpace (GNW) theory. He postulated that there is a centralized clearing house in the mind (the global neuronal workspace) which holds whatever stimulus the mind is currently giving attention to. Only one ‘stimulus’ (external, such as an attacker being transmitted through sensory receptors, or internal, such as a problem or worry) can be house in this workspace at any given time. Thus, the mind focuses its resources on the stimulus in the workspace until it is replaced with something else. Importantly, this means that whatever is in the GNW, and only what is in the GNW, will feel real. This means that whatever stimulus in contained in your workspace has conscious attention directed at it, and will feel like a truth in the world (even if it is not actually so).
Internal Family Systems, or IFS, is a model of cognition that views the mind not as a single entity but instead as a collection of subagents; numerous localized neural systems that are segregated from the rest of the mind (i.e. Friston’s Markov Blankets). These subagents each have their own dedicated purpose, unique optimization and decision criteria, and independent modes of processing. Generally speaking, a subagent will intake stimuli via sensory information or outputs from other subagents, and output its own package of processed information in the form or emotion or action impulse.
IFS casts subagents into two broad categories, Exiles and Protectors. Exiles are subagents that hold memories of past traumas and unpleasant sensations. Protectors are subagents that exist to keep the Exiles from entering the conscious mind (the GNW). In this model each subagent independently processes any given relevant stimulus or pattern of stimuli, and ‘votes’ on whether or not to let that stimuli into the GNW. When Exiles are introduced to the GNW, it causes intense emotions, often pain or fear. Thus, if a stimulus is triggering a traumatic Exile memory, the Protector responsible for monitoring that Exile may be able to override its vote, preventing the Exile from entering the workspace, and the associated negative emotion from becoming ‘real.’ If the stimulus is especially strong (e.g. if the stimulus is a bear charging your way), the Protector may be unable to stop the Exile from voting. Then the Exile’s output “RUN” will make its way to the GNW, feeling very real indeed.
These frameworks could explain things like why some people are scared of heights (the ‘heights’ Exile is particularly strong, so it regularly overpowers the Protector), why some people may feel unable to open up to love (the ’emotional damage and pain’ Protector is very strong, doing anything it takes to prevent that Exile from surfacing, even if that means sealing off emotional availability), and many other common and relatable phenomena. I think these models jointly provide an intuitive and powerful way of thinking about the mind, and point to methods for introspection and emotional awareness that I have personally found successful. While these are beyond the scope of this essay, I encourage you to read this series by Kaj Sotala for a thorough exploration.
Now we are equipped with a foundation from which to progress deeper into the mind. While Global Neuronal Workspace and Internal Family Systems frameworks are easy to understand, they function more as an intuition pump than rigorous model. They are highly anthropomorphized and leave large questions unanswered, such as:
- How do Exiles or Protectors actually ‘read’ stimuli and process it to create and output? What does output even mean?
- How does any output actually make its way into the GNW?
- Why don’t these subagents communicate with each other?
- We all know the mind changes over time as we learn new things and have new experiences. How does this ‘learning’ work with these systems?
These are all important questions and illustrate the shortcomings of highly abstracted models. However, they can be answered quite elegantly by implementing more detailed, lower-level models of cognition. These will be outlined and discussed in the following section.
Two mechanistic models of the mind
Karl Friston’s free energy principle succinctly states that to be alive is to minimize entropy. Entropy can be thought of as disorder in a physical sense, or errors in an information-theoretic sense. Friston’s concept has been widely applied to fields from finance to artificial intelligence to sociology.
In Surfing Uncertainty, Andy Clark, a neuroscientist and philosopher, applied the same free energy minimization model to perception and cognition by incorporating Bayesian probability. In summary, he postulates that perception is actually driven by two competing forces: One is the sensory stimuli that the body intakes (bottom-up processing), and the other is a model the brain generates based on assumptions built from past experience, known as priors (top-down processing). Clark calls this a predictive processing model of cognition. The brain generates a model of what “should” be perceived by the senses, based on priors relevant to a given situation. The goal of the brain is to minimize the prediction error of the models it generates. This can be done in two ways.
The first way to minimize prediction error is to make better predictions. Using Bayesian reasoning to determine the most likely aspects of a situation, the mind generates a prediction model. This prediction is then compared to the information the brain is receiving via sensory receptors. Any discrepancy is an error term, which the brain encodes and uses to update the prior assumptions by generating a new weighed conditional probability for a given prior under a given circumstance. The brain models the next unit of time using these updated conditional probabilities, fine-tuned to include the information from the error term in the last iteration. In this manner, the brain is continuously updating its models to minimize prediction errors, otherwise thought of as free energy, or entropy.
The other way to minimize prediction error to act on the world to make it closer to what was predicted. This concept is known as active interference, which simply means that the brain can interact with the world, altering the sensory inputs it receives. Consider the following example:
You are walking. Your brain’s predictive model for the right leg predicts something like ‘the leg will move in a repeating cyclical motion: move forward 3 feet by swinging from -30degrees to +15degrees from vertical. This generates a stream of sensory inputs: air flow on the leg hair, muscular contraction in the calf and quadricep, pressure on the ball of the foot. This cylce repeats approximately every two seconds, with the cycle ending when there is even pressure on the bottom of the foot.’
This model will have few errors generated as walking progresses normally. However, if you come across a hole, your brain expects to find a firm sensation under the foot to restart the cycle. Instead, it will find an unexpected error term – lack of any resistance beneath the foot. The brain then encodes this error term and propogates it up the model, decreasing the probability of “i am walking” and drastically increasing the probability of “i am falling.” This update generates a new model and set of predictions for the next unit of time, e.g. “my hands should be outstretched, i should feel the tension in my shoulders”.
Now, if the arms are not outstretched, there is a large error term. The brain can either revise its priors again to “i am no longer falling,” which would lead to a nasty head injury, or it can change the world-state via interaction, acting on the body to cause the arms to become outstretched, minimizing the error term.
To restate, the brain can minimize prediction error via updating priors, or by altering the data the sensory receptors are receiving by active interference. The former takes place when priors are weaker due to a relatively low Bayesian conditional probability (I thought i saw a bear, but it was a rock, so the prior was updated from ‘that shape is a bear’ to ‘that shape is a rock’). The latter occurs when the prior has a high conditional probability and updating could be quite risky (as is the case if you are falling on your face). In this situation the brain minimizes the error by acting on the physical world.
Predictive processing sheds much more light on what actually happens behind the mind’s veil, but alone still leaves many questions: what are these models actually made of, and how does this encoding and prior updating process actually work? These are still rather open questions, but scientists are getting closer to the answers.
Connectome-Specific Harmonic Waves
The brain is composed of neurons, which communicate with each other via neurotransmitters based on algorithms encoded in electric signals. This has long been known, and imaging techniques have been developed using levels of electric signal to measure brain activity (fMRI, for example).
Selen Atasoy has recently leveraged this fact to create her own neuroscience ontology, known as Connectome-Specific Harmonic Waves (CSHW). The defining discovery Atasoy and her team made was that brain waves travel across the surface of the cerebral cortex in harmonic motion with integer-based frequencies. The brain self-organizes itself around these unique equilibrium frequencies (known as Laplace eigenmodes). This means well-known mathematics long applied to thermodynamics and electromagnetism can be applied to the brain as well, providing a toolbox with which to rigorously test the theory quantitatively.
In short, the brain uses harmonic waves as a means of communication and computation, and is able to self-regulate these waves around various equilibrium points. This includes adjusting the frequency of waves to match external sources. I highly recommend reading Micheal Johnson’s Open Theory post on the topic, as he explains the nuances much better than I can.
Bringing it all together
A few disclaimers
We can now answer the questions posed throughout the preceding sections of this post. I will attempt to integrate aspects of each of these models, starting from the lowest level and building up to a holistic model of the mind.
As mentioned before, CSHW theory allows for the quantitative study of theories of the mind. The importance of this cannot be understated. Much of what I am about to propose is my own postulation, based on intuition and a bit of mathematics, but largely not experimentally verified. I will attempt to clearly note any claim that is based on empirical evidence to minimize confusion.
CHSW are the brain’s communication mechanisms, underpinning a predictive processing model of cognition
A 2018 study had Dutch subjects listen to speech segments of varying frequency. The initial frequency of the speech impacted the way subjects interpreted an ambiguous word at the end of the recording. While this may sound unimpressive, it unveils a critical relationship: there is a connection between the frequency of brain waves and what the brain perceives. There is a measurable cognitive impact of varying the frequency of brain waves. Otherwise stated, the frequency of brain waves impacts the Bayesian weights and conditional probability assigned to priors in a given mental model.
If true, this means that predictive processing models actually run on harmonic brain waves. I suspect these electromagnetic waves interact with the substrate by altering the the algorithm that governs a neuron’s interaction with others via neurotransmitters. Neurons generally have receptors for each type of neurotransmitter, and if a critical mass (the activation threshold) is reached, that neuron will fire and release an electrical impulse. By altering the way a neuron interacts with neurotransmitters, the harmonic waves change the activation threshold. I think adjusting the activation thresholds of specific neurons could be a mechanism for encoding conditional probabilities. Presumably, errors could be encoded and transmitted similarly, via waves of a different frequency adjusting the activation threshold again. This process would explain the bi-directional flow of information within the predictive processing framework.
Wave frequency-based communication also sheds some light on the ‘multiple subagent’ model of the mind. The subagents mentioned at the beginning of the essay are actually individual Bayesian models in the predictive processing framework. Exiles and Protectors are simply determined by the strength of conditionally weighted probabilities on the respective priors in each of these models. However, this does not inform us why these various models do not communicate with each other. I think that CSHW theory answers this as well: low frequency waves are able to travel farther distances, while high frequency waves are quickly absorbed (much like sound through walls).
Therefore, high frequency waves are likely used to convey local information and intra-model processing such as error propagation and updating of priors. Lower frequency waves would likely be used to transmit information across brain regions, such as when the output of a given model is strong enough to require conscious attention (e.g. send the information to the GNW). I think this explains why various models, or agents, do not communicate with each other. Communication between models really means computing joint probabilities with which to update both models’ priors. I would expect this is difficult due to the fact that high frequency waves are used for error propagation and model updating, and are inherently bad at travelling farther physical differences. If two models are not physically housed in adjacent bundles of neurons, such communication would be impossible. There is also (more speculative) reasoning that some neurons that compose models may only be responsive to waves of a given high frequency range, so they may be unable to use information conveyed via lower frequencies, even if it reached them. So each model operates independently, using high frequency waves to carry out computations until a lower frequency is needed to transmit across brain regions.
This lack of communication poses another problem to our holistic model: models jointly receive information from the senses, but do not jointly update. Different top-down predictions in each model mean errors will vary greatly, and active interference to minimize error in one model may generate increased error in another. This means that errors will ‘build’ up (commonly thought of as cognitive dissonance) across the models. We earlier defined predictive processing as an entropy-minimizing system, an optimization criterion incompatible with accumulating errors over time. The mind corrects this disparity using a very powerful mechanism known as neural annealing.
Annealing is a method used by metalworkers to strengthen metal. Metal is heated up to excite its atoms, which then begin to shift and move rapidly. As the metal cools, these atoms organize into more stable, stronger structures, increasing the strength of the metal. The process is similar in the mind. Sensory information enters the mind, generating error terms and increasing entropy. As the mind ‘heats’ up due to increasing entropy, the conditional probability on priors actually decreases as the neurons’ activation thresholds drop. This makes it easier for error terms to exceed the activation threshold and send high frequency waves out to update the prior’s conditional probability. On a large scale, this serves to ‘bulk update’ priors that govern certain aspects of the mind. Recent studies have shown a similar process taking place during meditation or psychedelic trips, perhaps explaining the peace some people tend to feel after these experiences. Scott Alexander outlines this process very well here.
Connectome-specific harmonic waves serve as the information processing mechanism of the mind. High frequency waves are used to power numerous Bayesian models that make predictions about the world, generating error terms when the predictions do not match sensory experience. These errors are propagated upwards in the model via high frequency waves as well, which alter the prior conditional probability by increasing or decreasing neurons’ activation thresholds. Lower frequency waves are used to transmit across further distances. These are likely the mechanism of active interference, carrying marching orders to various motor systems, or emotional responses to the limbic system. These models act as independent subagents, with the low frequency waves generated to activate neurons that move a stimulus into active consciousness, or the Global Neuronal Workspace. CSHW theory forms a basis from which one can integrate several neurological ontologies and create a more holistic, mechanistic, and quantitatively testable framework for cognition.