Visualizing ‘ensemble’ (Braitenberg, Schütz, Palm, Pulvermüller, Yuste, Buzsáki, Von Der Malsburg, etc.) neuronal nets. This is one of several rewrites and will not be the last. (An earlier installement: assembly-friends, runs in the browser).

My current line of thinking is to make something ‘biological’ (substrate-like), and put it together with a hyper dimensional computing - and associative memory - framework. (Current stuff).

The point of this is to explore how to bring in randomness (variation), selection (Neural Darwinism), conjecture and critism (David Deutsch) and meme-like (Dennett), self-growing (Turing, Levin), self-stabilizing (dare I mention autopoetic) software elements. (Such randomness is discussed in Mitchell and Hofstadter 1988).

Since I am a visual person and I want to feel what I program, I build this stuff here, to see what it does.

This is a kind of a-life excerice. And a hammock driven delopment thing.

My vision is a software world for Braitenberg vehicles. I suppose there is some law of nature that makes such software become a game engine.

This vibes with Josh Tenenbaum and collaborators Probilistic cognitive modeling.[fn:1]

Developing my personal lib for the visuals also makes this the place where I currently have the code for the softare art that I put up here.

When I close my eyes I am not blind. There is a shimmering there. Vaguely colorful blobs swirling. Perhaps manifesting into the hints of edges, then ebbing and flowing, washed away by some force of nature which they are not able to whithstand. Perhaps no more than the ideas of an idea of an object.

https://faster-than-light-memes.xyz/blerps.gif

• particle-field is a directed graph (on a 2d grid).
• Each node is connected to itself and it’s immediate neighbours (local).
• Time is discrete, at each time step A (activation) particles are selected (global inhibition model).
• This makes ensembles.
• They are in some ways a little bit like the gliders of this physics system.
• With adaptation (attenuation): Each node has a lower chance to fire when it is active (making it move like an ameaba).
• vacum-babble: Random elements fire (also intrinsic firing rate).
• decay: Random elements are erased (neuron failure rate).
• This gives ensembles a half-time.
• Activity must survive this decay assault, it must regenerate. (or die).
• the resonator part is the idea that top-down processes constrain which nodes are especially excitable for the blerp. (how to do this is to be figured out).

https://drive.google.com/file/d/1FzKIxnld6xk2b6WQIYcqZCyWtWSR4ii8/view?usp=sharing (does drive link work? Not sure).

[more soon]

This depends on ones sense of aesthetic.

Global inhibition and recurrent connections create neuronal ensembles. And the neuronal ensembles make sense to me as a mesoscale brain software element.

LLM Understanding: 2. F. PULVERMUELLER “Semantic grounding in brain-constrained neural networks”.

• emphasis on internally generated dynamics (Buzsáki)
• weights with log normal distribution, this makes ‘hubs’ / ‘backbones’ of activity
• classically, the recurrent neuronal nets (coming now in deep learning and cognitive modeling again as modern Hopfield nets) use local plasticity (for instance Hebbian Learning).
• I don’t need plasticity for the blerps at the moment
• There are other topics like spike rates, timings (The emergence of polychronization and feature binding in a spiking neural network model of the primate ventral visual system), synchronicity (e.g. Earl Miller) or computation at the dendrites.
• A brain network would have the dynamics to make stable trajectories (sequences) of hubs (ensembles, attractor basins), which are active for roughly 1 second. (Buzsáki The Brain from Inside Out).
• Since I want to model neuronal codes with hypervectors, I’m not sure if I need the trajectory dynamics
• Those are for the neuronal syntax, which is a rhythm encoding. But we don’t need a rhythm encoding, we need only get the essential concepts right.

• constrain the software to be biologically principled.

There are elements following local rules, which must self-organise, grow and regenerate.

Because we try to build software out of elements that must self organise, you get something that must be robust. It must deal with being unreliable and not-there anyway, because it must build itself in the first place.

I’m fine with trying to explore substrate-like elements embedded in a conventional software architecture. Using the power of computing that we already made available to ourselfes with software and programing languages, and tyring to build in some self-growing, self-exploring ‘subroutines’.

I.e. Something like blerps at the bottom, but perhaps just Lisp on mesoscale and macroscale levels.

Inspirations:

• Dave Ackley Robust First
• Turing patterns (Perhaps substrate-like computing is a continuation of Turings project).
• chemical computing

The neuronal memetics hypothesis one might say is that brain software is made from harmonic, agental (competent without comprehension) software animals.

I expect brain software to be something like ‘software that works more like music’. (Already the case on the system level: https://youtu.be/O4FCu1NqdYo?si=uZJP60ScDa9C-Uju).

I think memetics and neuronal darwnism is not utilized they way it could in systems neuroscience and neurophilosophy. (https://faster-than-light-memes.xyz/benjamin-overview.html).

The difference of toys and technology is nominal. I think that toys, games, user interfaces, computation and technology ultimately are on continua somewhere, but the same underlying theory will describe them.

I am a fan of pixars Inside Out (1+2), I love this idea of the mind as a colorful computer system. This colorful, joyful, toy-like information processsing system is the flair that I want my software art explorations to have.

Art diary: https://faster-than-light-memes.xyz/art-diary.html

• have some kind of python env

pip install torch numpy

I do this with

python -m venv venv
. activate.sh
pip install torch numpy

Then, I use run.sh via dev.el.

With the repl running, you eval a file like file:src/animismic/g.clj.

If you have cuda, this uses the gpu.

Buzsáki, G. (2019). The Brain From Inside Out. New York: Oxford University.

FRIEDEMANN PULVERMÜLLER https://www.sciencedirect.com/science/article/abs/pii/0149763495000682

[fn:1]

I see value in this, but my fundamental idea is that a cognitive system makes it’s own physics. A world model is a useful technology for the cognitive system running an animal in the world;

But my claim is on the nature of the software, not what it computes. Exploring possible worlds and possible physics should be equivalent to program search.

In other words, there is a kind of physics and chemistry that the brain is implementing, and that this makes a simulated virtual world in which software-biological entities (memes, subprograms, concepts, habbits, mental-technologies) live.

What is this physics and chemisty? That is the topic of neuronal codes, syntax and interpreters. (And I’m trying to make toys that explore them).