Make a system for signals like in godot. Functions call a queue for globals variables, every variable is a pointer to another function to a sum od arguments which was defined by a queue of variables set by a set of functions. A server collects the current refrence of each neuron to every of it's consequential neurons, then updates all neurons accordingly. Transfered is an integer modified by a number of methods inside a neuron object. There must be a way to generate neurons such that they integrate in an alreagy operating system based on that systems state. It might be less efficient then suffeling neat tensors, but it might yield an agent that learns from experience.Every cycle each neuron calls a number of neurons in a set of which it is a part of based how many neurons called it prior. A function taking the number of calls as an argument generates a list of neurons to be called next cycle.example 1 calls 2, 4, 72 calls 1, 23 calls 5, 7, 8, 9...N calls some number X of N's based on how many N's it was called by Y. The number is a non linar function of Y(X), based on Y a list of neurons is generated dynamically. Instead of neurons having variable weights which connect to every single neuron in the next layer the varibale determines how many neurons are called by each neuron, with every weight eother being 1 or 0. That way the system could add neurons and connections during operation. How it learnes is determined by the way neurons are added and coneccted to other neurons. The fine tuning of this algorythm can be done via evolution. So it can add neurons as it operates but how it adds (or subtracts) neurons despends on a preprogrammed algorythm. This algorythm must organically interlock varying types of connection patterns which are flexible and react to the current state of the whole system. Single bits instead of bytes as weights could reduce complexity by adding neurons. Possibly this could be made more efficient by programming the system on a FPGA.
>>>/x/
Create multiple types of neurons with various methods for connecting to other neurons. As networks of higher complexity are created they might organically form functional clusters. Neuons can be added or subtracted based on activity. Instead of connecting neuons to all other neurons the nurmer of neurons is limited such that complecity doesn't increace exponentially with neuron count. A virtual enviorment might select for systems with functionallity once enough degrees of freedom are added, kind of like a simulated microcosm where we can observe how different systens if growing neurons perform.
>>109942162gtfo I made an algorithm, not superstitious in any way
clusters clould inherit methods based on pretained systems, systems could be grouped in blocks to reduce trainibg time and allow for more effective scaling.sorry for typos, I'm extremely sleep deprived
>>109942134Untrustworthy handwriting; opinion disregarded
in a classical neural net neurons in each layer connect to all neurons in the next layer and layers dont circle back or skip any layer. there should be just one pool and with each neurons having the potebtial to connect to any other given neuron but only connecting a limited count to reduce complexity. the goal is to give the algorithm the ability to process and learn while it processes. it's not a ready made model that was trained prior but a system that organically changes in form and size based on input.
>>109942323what does untrustworthy handwriting mean
cowards
to find a system which connects newly generated neurons to old neurons and subtracts neurons based on how the current state of the system is the key to making an AI model with generalized abity to adapt to bew inputs. It will delegate resources to neuron creation where needed, thus providing enough flexibility to adapt while operational. how else is AGI supposed to function? axons send either 1 or -1 as weird, all weights passed to a neuron are added to determine number of neurons called in the next cycle. the specific neurons dont have to overlap for various counts. instead they could stretch or disperse based on various operations. type a calls x^2, type b e^x etc making building blocks for organically interlocking clusters which could be added dynamically.
holy fuck I need some sleep asap, feel freebto add ideas on how to make an algorithm like that
>>109942355NTA but looks like what someone experiencing a psychosis would writeNot the content...well kinda...just the shape of the lettersCompare it to, let's say, XV century letters from the aristocracy and you will notice your way of writing looks unhinged
>>109942747nah I took my meds, I'm tweeking though
why don't you write the code?
>>109942134I feel like the A (in pink) and you are the B (in blue)
>>109942134If you're trying to a continuous learning algorithm, the most important parts turn out to be how to alter the strengths of connections based on activity, and how to forget stuff so that the network doesn't get bogged down by irrelevant shit. If you have those, you can get away with random connection formation (provided you have enough scale). It's not quite enough for AGI (that requires some nodes to do XOR between inputs in some sense) but it is enough for pretty good continuous learning. And congratulations, you're only about 10 years behind the cutting edge, unlike the goddamn "AI labs" who are decades off the science.If you were doing recurrent networks, I'd note that you need some nodes dedicated to stopping the network from going into spasm through inputs with negative weights. You need negative feedback as well as positive feedback. The exact same learning model can be used with both types of connections.
