I recalled that there are sensory topographic maps in the brain. Why would the brain do that, unless it wanted to apply a sequence of not too drastic geometric mapping to the data? Perhaps along the lines of linear mappings.
https://archive.org/details/cortical_geometry_ccslm
The note should have had a more speculative tone TBH but it is what it is.
I devised this story:
https://archive.org/details/sherlock-holmes-and-the-astonishingly-simple-neural-network
Maybe it can help provide you with some astonishment.
It also suggests that neural network researchers have been applying recursive dismissal to their own subject, at the foundational level.
Here is an image of recursive dismissal in neural network research:
You may be making this too complicated. The strips are where the sensor data lands. Topology carries important information about the sensed data and evolution has shaped the processing to exploit this fact.
На одном из ваших диалогов, мне очень хотелось выразить свое недовольство. Вы спекулятивно используете интуитивные выводы поддержанные за последние 100 лет. Когда это вижу, меня это переворачивает. Поверьте, я не хочу быть вашим противником, потому что вы меня просто исключите от форума. Но я Вам заявлю прямо. Вы во многом правы, но интуитивно ошибаетесь. Вы консерватор, и используете уловки и приемы которым сложно противостоять обычному пользователю. Но я вижу эти уловки. Я вижу их в книгах почтенных. Давайте сформируем диалог. Кажется что я, что вы находимся в неком информационном пузыре. Я предлагаю выйти из зоны комфорта и потягаться со всей Российской классической наукой в лице меня.
The brain didn’t “do that”. The logic of ‘why’ and ‘unless’ and so on does not apply.
There are sensory fibres from distant sense organs through spinal column and brain stem to cortex, which evolved from small simple organisms to the complexity we see today.
It’s reasonable that fibres from one finger will lie alongside fibres from the adjoining finger, because of a lack of evolutionary advantage leading to a different outcome.
Random projections and random wiring (if enough of it) can fully preserve information despite you nor I being able to make head or tail of it. Why go to the effort to deliver everything in pristine geometric order if there is no informational purpose to that evolutionary fight against entropy?
It is not trying to deliver everything in pristine order. The fact that geometrically two close sensory sources (e.g. two fingers) end up being connected to relatively close areas in the brain is simply the consequence of how bodies grow, no for a special effort to achieve that. And yes some randomness does happens but growing in total randomness (disregarding any tissue geometry coherence) is what cancers do - resulting in a dysfunctional body unable to survive.
Even if information is preserved learning might be more difficult when one finger sends a sensation of tickling while touching a near area delivers a flowery smell. I mean it’s possible with a great effort to make sense of it all, but getting two close representations when two similar events occur helps.
And finally chaotic connecting needs chaotic growth, that might be more difficult to make right since some hyper-parameters still need to be controlled.
Random does not preserve information. Effort, pristine, purpose, fight: these are words that do not belong in any argument based on science. If you have a point to make, please keep it scientific.
Neurological structures perform a useful function, even if we don’t know how. Every organism grows from a few cells to many; lives, eats, breeds, dies as its parents did before and its offspring will do afterwards. A structure inherited from parents is passed on to offspring unless there is an advantage in doing otherwise, and even then each difference will be tiny.
Every organism (including your good self) has structural features inherited from millions of generations back (eg notochord), because there was no evolutionary advantage in doing otherwise.
You can apply a permutation to an image and lose zero information. It’s just a question of whether an inverse exists. The permuted image looks like garbage but every single bit of image information is still encoded within. If you want your nice geometry back apply the inverse permutation.
Likewise with invertable random projections.
Humans have a big problem with assuming things.
The information from sensors could be supplied to the brain at the convenience of evolution in so to speak permuted form. That it is not is probably saying something. Why make the extra effort to deliver it in non-permuted form?