Thursday, 13 March 2014

The Convoluted Mind: How Experiences Distort Our View Of Reality

As a hobbyist AI researcher I have often found myself pondering on the exact workings of the human mind, and biological neural networks in general. The most fascinating aspect about it I have always found that of the creation of what I have come to call an 'internal universe', or IU. Essentially it's an internal representation of reality which doesn't necessarily match up with the actual reality. This IU can also contain completely imaginary things which in actuality do not exist, or distortions of factual processes to give it another interpretation.

Before I dive into this topic of internal universes, though, allow me to first address the nature of a biological neural network (BNN) and the way experiences and memories affect it. The essential aspect of a BNN and neural networks (NN) in general is that of input. Without input (the 'brain in a jar' example) there's no point to an NN. The basic and most essential function of an NN is to transform or convolute input. Lack of input will generally lead to self-destruction in NNs with active internal feedback mechanisms (see sensory deprivation experiments).

This input we can refer to as 'experiences' and the recollection of such input as 'memories'. They form the foundation and corner stones of the functioning of a BNN. Their role in the BNN is both formative and functional, in that the input (I) and recollection (R) are required during the formation of a BNN to form the proper structures. This process can be observed in infants, for example, where the type, intensity and duration of input they are exposed to during their first months and years can lead to the formation of a BNN which produces very distinct and quite predictable output (O).

A NN is by definition very unlikely to respond by itself in a manner which can be considered logical. Instead it appears to become 'programmed' by the input, which modifies and adds to the convolutions applied to the input signal. Most of the resulting behaviour (output) is simply handled by fixed parts of the network, resulting in what is referred to as 'instinct' or behaviour routines. This results in a predictable, fixed response (convolution, or C) to a certain type of input, in the form of I -> C -> O.

Naturally, the intriguing and defining part of a NN is its self-modifying functionality, which is of varying capability in different individuals. This process takes place in different parts of the BNN, from the lower, less advanced sections of the network, to the newer, more complex sections. In the former sections we see the above process, of input directly modifying the structure of the network. In the latter sections we encounter what is commonly referred to as 'intelligence' or 'reasoning', but can also be called more broadly as self-awareness (SA).

SA is the ability to reflect upon and predict the consequences of a course of action. It is a reasoning process involving the awareness of oneself and also the essence of what AI researchers are trying to emulate. SA is also the core of the IU. Imagine SA at the core of this universe, with recollections, input, convolutions and active feedback (AF) surrounding it. This model is essentially the model of a personality, as it's referred to. In this model, SA is a semi-passive presence in the NN, monitoring its surroundings and trying to compose a functional model of the reality outside the NN. This leads to the IU.

In short-hand form: C(I + R + AF) -> SA(IU) -> O.

The interesting thing about this model is that it seems to form an adequate model for human behaviour and interactions, including in large groups. It explains why the old adage that no two people will ever fully agree on anything is theoretically true, as the possibility of the IUs of two individuals matching up fully is extremely remotely. While the SA can semi-directly access the I by suppressing convolutions and thus validate the IU against reality (the underlying principle of the scientific method), this is a limited and intensive process. Only a strong SA can sufficiently do so, and may then still be blocked by the limitations of the IU due to previous input. This is the underlying principle behind psychological trauma, where a sufficiently strong input can cause a strong AF process involving R. This is what happens for example in PTSD, where the AF result can be overpowering, even for the SA and its IU.

Soon I hope to further test this model using software and FPGA-based implementations. The prospect of this theorem making it into a replacement for psychology and foundation for AI seems promising at this point.


Maya

3 comments:

Manuel Dornbusch said...

Good thoughts

I like to remind people that our brains, in a simplified manner of representation, are nothing more than very very good pattern recognition machines.
Each process of learning or applying learnt information is obviously just recognizing, storing or repeating patterns.
We are so good at pattern recognition that we see patterns, where there are none. Sheeps and bunnies in clouds.
To go to your model, your SA works with wrong data, because the I already goes through the pattern recognition Black Box. SA does not judge the I on raw data. The data from I is already interpreted, or filtered.
I agree on your last paragraph as it covers mostly what I explain with the problems of the communication model.
Most people see this mode
Sender - Receiver
But both sender and receiver filter their O and I by filters that are formed through a lifetime by experiences (recognized patterns) What I say in any form of language is a poor representation of my thoughts and the same words, as the arrive to you, mean slightly different things to you. Add to that more filters, like social layers (how do I want to express myself to appear to you in a form that I wish to appear to you; what is your general attitude to I coming from me, by your social view of my person, your experience with me...)
...it goes just further down the rabbit hole.
To go back to your model. Two IUs will never be able to exchange data without error, because the communication is flawed through filtering by applying different learned patterns

Maya Posch said...

Well put, Manuel :)

Another thing I didn't really touch upon in the article is that of feedback loops between IUs, which tends to lead to strife over a completely imaginary topic. See countries, economies, religions and wars as examples of this. It's essentially the selfish meme, as written about before by others.

The idea that a thought or concept can propagate itself throughout a population like a virus, multiplying and evolving, but ultimately doomed to vanish again in the brutal meme ecosystem.

Manuel Dornbusch said...

How does this fit into your theory:

If SA expects an input (because this is a previous recognized pattern) but does not get it, or only incomplete, it will add false data to fill in for the blank. The NN has no authenticity control.

Everybody has false memories

Example:
Most of our communication these days is written. Meatspace meetings or at least voice/video streams take a far behind second and third place.
In communication with friends in the USA I fairly often noticed misunderstandings. By analyzing the situation with a friend in America (with background in psychology), my grasp of the English language was not the problem.
Each time I meet these persons, or when we video chat, any previous misunderstandings could be cleared and no new came up. Many people underestimate the nonverbal communication channel that is completely missing in written communication.
So when that whole communication channel is missing, the brain fills in some of the expected, but missing data. What is filled in depends on what you said: memes, nationality, history, any form of group identity