The 'wiki' section of the site is the progress of being migrated from an old site. There are some rendering issues and many of the links don't work yet.

Background to A Theory of Mind

This article investigates how consciousness in the human mind operates, explaining it from common phenomena. The focus is on providing an explanation for how I drew conclusions on the architecture of consciousness.

For further reading see [[A Theory of Consciousness]], which provides a unified description of consciousness.

In order to explain why consciousness is what it is, you will be taken on a journey through your own mind. This journey will attempt to peel away the layers of obscurity, and to drill down through uncertainty when there are things that we cannot directly measure.

A Mental Exercise

To begin, you will need a stimulating drink. So please take 5 minutes to make yourself a coffee or tea, or other drink of your choice. While you are doing that, please observe your thoughts. There are things to discuss upon your return.

Making Tea

The act of making tea involves many steps. Some of those steps can be done at different times, while others require a definite order. Some of those steps have some flexibility, while others must be carried out exactly. If you were to pour the coffee grounds into the cup, instead of the plunger, you will have a very grainy coffee. Or if you were to pour the made coffee, before getting a cup, you will have your drink all over the table and floor. You probably didn’t notice many acts of conscious decision making happening during the process, and you most likely didn’t notice each of the individual steps. You almost certainly did not notice any effort required to plan the stages, their order, and their timing.

Assuming you chose to make a drink that you are familiar with, you largely carried out the tasks in auto-pilot, because you already know each of the steps too well to require any conscious thought.

Trying Something New

That task was too easy. Time for something more challenging. You need to try something you don’t already know how to do. Thankfully the internet is full of brain teasers. And they’re perfect for the next exercise.

Please take the next 5 to 15 minutes to find a couple of brain teasers you have never done before, and try to solve them. While you are solving each brain teaser, pay careful attention to your thought process. In particular:

In your thinking process, how do you decide to move from one step to the next?
How do you know what the next thing to try is?
When there are multiple things you could try, how do you decide which to try first?

(tbd: provide a list of good example brain teasers)

Here’s the big thing that you will hopefully notice during that exercise:

Some thoughts just appear.

Spontaneous Thoughts

This is the first glimpse into how the brain functions. In common language, we call this the sub-conscious. It’s full of learned knowledge, mental skills, heuristics, mental short-cuts, and strategies for blind guessing. It’s extremely adept at taking a problem, and throwing suggestions at it. Some of those suggestions will be correct solutions, while others will be completely unhelpful. As we develop and mature, the sub-conscious gets more and more accurate for our common everyday tasks.

In the terminology of Danial Kahneman this is System 1 or Fast Thinking.

We can do a lot with only System 1 thinking. In fact, we spend a lot of time doing only System 1 thinking. Keep that in mind, as we will come back to it shortly.

Neurons in Space

In software AI, Neural Networks are typically stored as data. That makes it easy to duplicate them, move them, and to apply them to some other inputs that came from a different source than where they were used previously. The physical neurons in a brain do not move. If you want to process a different piece of input data, using the same neural pathway as before, you need to move the input data to the neurons.

Likewise, our brains are full of many different individual regions of neural network that each focus on a particular task. If we want to think about something from our past, we cannot direct the appropriate neural network region to act against data directly in memory. Rather, that memory must be loaded into a specific location in the brain that acts as the capture point for that region of neural network. But, having a different capture point for each task-focused region would lead to very limited thought patterns.

It’s more likely that there is a single region within the brain, that holds the data being actively worked on, enabling many regions of the brain to observe and act against that data at once. This is the purpose of working memory.

Task-Focused Processors

If much of the brain is divided into regions that focus on a particular task, and they act only on the data within working memory, then it seems likely that each individual task-focused region operates in a System 1 Thinking way.

Referring now to those task-focused regions attached to working memory as Processors, here’s a model that they could operate on:

Some input is loaded into working memory
All attached processors have the potential for producing an output.
Some automated selection process is used to pick the best output. This is likely some sort of competition mechanism, such as picking the outcome with the strongest signal strength.
The selected output is loaded back into working memory as the new state, and the process repeats.

The description here seems remarkably like a generic computing engine, something akin to a Turing Machine.

There is more to the puzzle yet, but this is one of the more significant points here:

the consciously aware part of our brain is a biological computing engine.

Awareness of Thought and the Error Prone Brain

In the exercises above, you observed the flow of thoughts while you carried out actions and solved a difficult problem. But why can you observe them at all?

Consciousness as a Sense

tbd

Error Prone Brain

The high reasoning part of our brain is a state machine, but without the reliable stack of a computer CPU. The architecture of having many processors competing for attention is extremely flexible, but suffers from a susceptibility to wandering thought, circular thought (infinite loops), and stuck thought (dead ends in the state machine).

According to my theory, basic creatures like insects probably employ purely stateless processors. The only state coming from body feedback and hormones. So the evolution of higher reasoning also needed to develop a mechanism for managing the overall processing. That management process is complex, and needs domain knowledge of the data and thought process being carried out. So a natural solution is to just feed the mental state break in as a first class input. Interestingly, this would lead to evolved and learned processors with the processing of problem data and feedback being tightly integrated.

This doesn’t need general intelligence. It’s a fundamental architecture. It doesn’t need tight integration of problem and feedback data, but it helps. It does need a sufficient level of intelligence applied to that feedback, and if there’s no general intelligence, then it basically has to be hard coded.

So there’s basically two options for achieving that level of feedback understanding:

Hard coded feedback processing and attention control. Limited data integration. Doesn’t require general intelligence. Very unhuman-like.
Fully integrated General intelligence with learning. Human-like.

In the human case, we are probably born with some basic building blocks of attention control - that do only the most basic of attenuation of repeated signals to avoid tight infinite loops. And then we learn to override those and apply higher reasoning.

This theory also suggests there is a minimum level of intelligence required for consciousness, because otherwise it cannot control thought well enough. But in contrast, maybe that limit is very small…the slightest bit of self control could rate on the consciousness scale.

A little example:

If you combine multiple unreliable components together, such that their errors are multiplied, you very quickly converge towards zero accuracy. But if you verify one component against another, you can increase the overall accuracy. The likelihood of two 90% accurate components both being wrong is only 1%.
tbd: move this to discussion on background

A Necessary Component of the Architecture

So the awareness of our own thoughts is a fundamentally necessary input to our thought process, in order to govern its activity. And the integration of working memory and processors now looks like this:

Some input is loaded into working memory
All attached processors have the potential for producing an output.
A learned selection process is used to pick the best output.
The selected output is loaded back into working memory as the new state, and the process repeats.
The short and long-term outcomes are monitored and critiqued over time, and controls are employed wherever necessary to adjust how the thought process is carried out.

Processor Output Selection

The attention control mechanism monitors the quality of results from processors, and attenuates their strength according to that quality. Over time, as processors are trained, their accuracy improves and their strength is increased.

In reality, the biological architecture of processor output selection is probably quite different. For starters, we observe using fRMI that certain inputs seem to only activate related areas of the brain. Additionally, neurons work not on a continuous output state, but on electrical spike firing rates. So it seems reasonable to conclude that processing of a sense spans out like a wave, and is filtered for at each step, so that eventually only the final selected processor completes the execution of the wave back into the attention centre and into working memory.

Experience of Consciousness

One of the greatest unsolved problems of consciousness is why we experience it in the way that we do.

There are a few aspects to that:

Awareness of Thought as a first-class sense.
Mental Model of Mind
Awareness of Thought integrated
The thing that experiences

…tbd…

Scope of Awareness

One fascination outcome of the theory above is in how we are aware of some thinking processes, but not others….tbd…

tbd: size of working memory vs CF.

The odd selection of catchment area is likely an artifact of what CF is needed for : to manage errors.

Learning

Fitness functions - minimise surprise.

Why is individual intelligence limited?

The architecture I have theorised does not imply any limits. I believe the same architecture is used throughout all mammals at the least, to varying scales. But clearly the DNA of a species puts limits on the capacity of an individual to increase their intelligence. But why? Why is it that a dog cannot learn to be as intelligent as a human?

An answer to this may lie in the fundamental difference between what is controlled by DNA and what is learned.

DNA must set some controls on regions of the brain. The visual system presumably cannot be re-purposed to rationalise about thermal systems. And that makes sense - the visual processing system is architected with a focus on visual inputs…so it’s learning is focused on that.

Additionally, learning requires a fitness measure, which will likely be heavily controlled by DNA with little, if any, adjustments possible from life experiences. The fitness measure of each region is brain will be heavily focused on a particular task. This likely explains why the brain seems so conveniently modular - learning is more efficient when focused on one thing.

So the limits of an individual’s intelligence growth are imposed by the specific focused tasks that evolution has selected for.

A concrete example may actually lie in the way that we can increase intelligence to an extent. Individual regions of brain can increase in size. For example London cabbies, with their enlarged memory centre. But it is likely that our biology imposes limits on that growth. Any fitness function will incorporate efficiency, and there is an inherent trade-off between size and efficiency. The limit is “soft”, and can be stretched, but only so far.

Beyond that, each brain region had an inherent capacity. This is like a (roughly) fixed number of neurons that can learn anything its given, but only to the resolution possible within that number of neurons.

Evolution

Consciousness

We have discussed above how consciousness is a necessary component of our brain as a computing architecture. I believe that alone explains why consciousness evolved. As we started to develop high-level thought, we needed a control mechanism. Otherwise we would have ended up with a system that is highly prone to pointless wandering thought, with long wasteful cycles.

Evolution would have created many different variations on architecture. For example:

Different scopes of awareness - ie: different parts of our brain being included in the catchment of conscious feedback.
Different levels of course-grained high-level thought vs fine-grained low-level details within conscious feedback.
The feedback coming into the system at different places of the architecture.
Different mechanisms for how thoughts can control the action of the processing architecture.

Ultimately, the design we have today will be based on what turned out to be the most efficient and effective. For example:

Scientists believe our working memory can hold multiple things at once (approximately 7 items), and yet we are only ever consciously aware of one thing at a time. I believe this means that the conscious feedback provides access only to the single top-most (or most recent) item in working memory. This must have happened because it turned out to be sufficient to maintain stable thought.

Fitness Functions

tbd

Mammals

All mammals have basically the same brain structure. In particular, all mammals have a neo-cortex, which is thought to be the primary component involved in high-level thought. So it seems reasonable to assume that all mammals are conscious.

Other creatures, like birds, fish, and insects, do not have a neo-cortex. So perhaps they are not conscious, or certainly not in the same way that mammals are. Remember, consciousness is an architecture, and it’s there to solve a particular problem of control. If these creates have a different architecture, then they definitely have a different experience. It’s possible they have some form of consciousness too, based on a different architecture.

I believe animals can function without high-level thought and thus without the need for consciousness, so my best guess is that non-mammals do not have consciousness.