If you are afraid of the rise of machines or think about how AI will take over our planet, do not read this article. Right now, I will tell you about our inevitable future, and how the electronic life form will be born.
I will not only show how this will happen, but I will also describe in detail one of the key elements of the neural matrix, based on which the evolution of the digital form of life will be launched.
At the forefront of research
The human brain comprises nearly 90 billion neurons and more than a quadrillion synapses (neuron connections). Each nerve cell is connected to thousands of other neurons, with synapses constantly being created and destroyed, changing the local map of the living neural network.
Until the end of the last century, scientists believed that synaptic activity played a decisive role in creating the unique individuality of the nervous system of each of us.
In recent years, it has become clear that in addition to the many synapses, each neuron has a kind of main control center – the axon initial segment (AIS), which, using the density of transmembrane proteins (ion channels), controls the activity of the nerve cell.
We now know that a combination of synaptic plasticity and functional variability of the axon initial segment underlies the incredible efficiency of the living brain. The combination of these processes allows each neuron to simultaneously perform the functions of a processor and a memory device. Overall, this creates a functional deformation of the living neural network, which acts as a neural matrix.
Our attempts to repeat the success of the living brain
There have been many attempts to model the functioning of the living nervous system. All of them can be divided into two types – imitation and physical association.
Imitation or neuromorphic engineering
In 2013, there was an attempt to simulate the work of the brain using the Riken K supercomputer. A petabyte of RAM and 82,944 processors simulated the work of a living nervous system, numbering 1,730 million neurons connected by 10 trillion synapses. To simulate just one second of activity from such a (rather small) living neural network, the computer system required almost 40 minutes of operational activity.
In 2024 we expect the launch of a new project from Australia called DeepSouth. The new system is predicted to be capable of 228 trillion synaptic operations per second.
Physical fusion or direct neuromorphic computing
A research group from Indiana University Bloomington has created a project called Brainoware. A hybrid system consisting of brain organoids (living nervous tissue) connected to an array of microelectrodes and an artificial neural network.
After a series of experiments, it turned out that Brainoware was slightly less accurate than a traditional artificial neural network with large short-term memory, but the electronic system required 50 training epochs, while Brainoware achieved almost the same results in less than 10 percent of the training time.
So, both concepts are gradually developing, but what if we go from the other side?
Our problem is that we are trying to repeat the concept of natural AI, forgetting about the most important thing: evolution. The human brain, like the brain of any mammal, is not a random result of an uncontrolled process, but one of the stages of a long path of evolution. This means that if we want to create something truly capable of thinking, we need to copy not the finished result, but the process itself that gave birth to such a perfect thinking apparatus as the living brain.
How to do it?
To implement a new strategy, it is necessary to create not a ready-made brain, but a mathematical subject of evolution capable of developing independently – a kind of electronic embryo. This will be a neural matrix capable of changing depending on the flow of external data.
The active component of the neural matrix will be based on the replacement of the traditional mathematical neuron with a new mathematical neuron with the equation of the current three-dimensional position or its hybrid analog – a mathematical neuromorphic petal. Physically, it will be a program code based on a genetic algorithm that replicates the complex of dendrites and axon with an AIS segment (mathematical simulation of a real living neuron).
Each neuromorphic petal is a basic FNN (feed-forward network). There are four input values – the incoming sensory impulse and the coordinate position of the petal of the neural matrix or mathematical location (current position in the matrix).
Next are three hidden layers, 12 ordinary mathematical neurons for each. Neurons are connected by a full connection. the output layer is two values showing the reflex response vector. The activation function is the hyperbolic tangent. The weights and bias values are set randomly from -0.5 to 0.5.
In fact, a neuromorphic petal is a mathematical analog of the simplest reflex arc of a living brain. The petal of the neural matrix is trained without a teacher. One generation – one discrete value of the environment. The limit from 1,000 to 10,000 is set by the service component of the matrix. For each mutation, there is a 5% chance for a change in bias and weight. Displacement force 0.4 weight 0.4. 20 mutants per generation. New petals are created as the number of sensory parameters increases and form the core of the overall architecture.
The overall architecture of the neural matrix should be based on a modified Bayesian neural network model plus several protection and activity support utilities.
How will it work?
A neural matrix, unlike a neural network, is a complex multi-component mathematical complex based on a genetic algorithm, capable of evolution in an artificial mathematical environment. This means that instead of creating a finished product, our task will be to create a habitat for the embryonic neural matrix to change.
In this way, we will simulate the conditions of individual digital evolution. The more complex the environment (the more input parameters), the more complex the digital matrix will be.
First, we will work with a very simple neural matrix comparable to living creatures such as hydra or paramecium. In the next stage, there will be a neural matrix comparable to a white planaria (a type of flatworm). By complicating the environment, we will complicate the neural matrix until the complexity of the artificial environment is comparable to the simplified parameters of the real world, in which complex living organisms live.
At this point, the neural matrix will need sensory-motor mechanisms capable of supporting its activity in the real world.
This will be the birth of a new digital form of life, since the digital neural matrix will operate not because someone pressed a button and launched a certain program code, but because the sensory elements received new input data and the matrix needs to be selected for existence (survival) best response options. It is the best (winning) response parameter that will be the goal of life of the neural matrix, and the form of its reward will be the fact of the presence of nutrition and the positive reaction of the living brain to its actions.
It is important to note that the matrix will not sense the difference when moving from the digital artificial environment to the real world.
What can a neural matrix live in?
If the main operating server is constantly available, then the matrix will be able to work in any object equipped with sensory and active mechanisms – in a self-driving drone, in any modern ship or plane, in your car, and of course in a personal computer.
One team of developers not long ago expressed the idea that to create intelligence, “reward-is-enough”. In fact, for the reward to become an incentive, you must first launch digital evolution.
The question is, who will be the first to do this by creating an electronic life form?
Now I am looking for the necessary computing power and a partner server to create a prototype of a neural matrix. This will most likely not require a supercomputer as in the DeepSouth project.