AI - computer vs biological logic
Computers
Today, everyone talks about artificial intelligence and assumes that it represents the pinnacle of our technological development. The theories of parallel processing, workflows, and artificial intelligence have yet to be developed. The large number of related sciences on these topics makes it very difficult to reach agreement to meet the requirements. In fact, computers are still the basic prerequisite for AI. Regardless of the type of application we want to implement, the von Neumann architecture is always used in conjunction with programming languages, all based on Ada Lovelace's concept of algorithms. Let us have a closer look and use some comparisons for a better understanding.
Imagine the following: a floor and an artist with a toolbox. Now this person is looking at the floor, while at the same time nothing is moving. The plan of how to change this floor is irrelevant, what is important is that only one person can make changes, and the changes can only be made with tools from the toolbox. Also, the floor is divided into small areas, so we have a mosaic. If the artist wants to change something, he must know exactly in which quadrant he will do it. The plan is to have a certain image or many moving images in a certain time.
So let's look at how he does his work. First he takes the corresponding piece in his hand and changes it with the necessary tools, then he puts it back to its original place or maybe to another one. Well, the artist corresponds to the processor, while the mosaic represents the memory. Since there is only one artist, the work is done serially, one step at a time. The toolbox corresponds to the instruction set of the processor. Sometimes the artist has a small rack next to him where he can put the several pieces of the mosaic before putting them back on the floor. This corresponds to a cache, that is, a small temporary memory.
All the steps that need to be executed, we call software. Sometimes, and more often than we would like, this software is very difficult to create, so we make it easy and use programming languages. These are a superset where we combine some initial processor instructions into a logical step. We also have some predefined software in the form of libraries that we can reuse in certain circumstances, this also makes programming easier. The order how to execute a task is broken into many small steps, where the dependency is regulated by advising which is the next step to be executed. This is what we call an algorithm.
If the task requires too many steps, we can increase the number of steps executed by increasing the working frequency. In fact, the processor more or less always works at the highest frequency. If there is no task to execute, then the processor does nothing, namely in a loop until it receives a new task. Since the operating system is user-driven, that is, the external user defines the required tasks, this means that most of the time the processor does nothing and just waits for new tasks.
First there is a plan and a piece of "something", as a matrix of static pieces that we used to call a mosaic, that we want to change or adjust to meet the requirements. Then we have a worker who does the tasks step by step. At the end, we have the tasks that need to be executed. These are computers and the above described approach is called top-down.
Biological logic
But there is also a bottom-up approach, which is used in biological logic, aka the brain.
The architecture of the brain has a distinct hierarchy, the 1st level is built from neurons through groups to neural networks, the 2nd level takes the neural networks as building blocks to implement various functions and create higher forms of organization by grouping them into a specific landscape to enable a knowledge domain, while the 3rd level takes the domains to connect them into a region with specific purpose while grouping them as needed to communicate with the other body components such as sensors, actuators, and internal organs.
Recall the mosaic comparison from the computer world. The biological equivalent of the building block of the mosaic is the neuron. Moreover, there is neither an external artist who can control all the others, nor a toolbox with special functions.
However, the neuron is not static, but dynamic, even interactive. And why? Because it is not a stone, but a living being, a cell. Each neuron must manage its own capabilities, which requires it to join with others to form a group with a specific function.
?However, since it is very difficult to imagine such architectures, I will use another example as a substitute. Imagine neurons that are all independent but live together like a colony of bees. That's better, but still not good enough. How about we take humans as the building block to create a society. That's easier to understand and also easier to come up with the numbers we need, so we can think of the brain like a city.
However, this is not entirely true because the capacity of the biological brain varies by species. For example, a honeybee has a brain with about 1 million neurons, which is equivalent to a city with 1 million people. The mammals/fish/birds have a brain capacity of at least 1 G neurons, which is equivalent to 1 billion people or the human population of a continent. Primate brains have at least 10 times that capacity, which is equivalent to the entire human population on Earth.
领英推荐
So let's start with some small numbers. For example, a neural network has 4K neurons, which is equivalent to a village or a small town. How do people in this town work and organize themselves? Well, first they form a group for a specific purpose, then other groups with different goals. But wait, the same person belongs to more than one group, so we have more connections and interactions between one person and other people.
Most groups are about accumulating "things", others are about changing things, and still others are about managing all the other groups. So we have 3 types of groups, where the number of accumulators and changers may vary, but there is only one group that manages all the other groups. Et voila, we have a network of people that follows the same rules as a neural network.
Now we can expand this even further by creating a region with many cities, where each city has its own specialty, but together they have their own specialty knowledge, then we have countries formed from these regions, and in the end we have the continents and all of humanity on Earth.
This sounds complicated and it is, but this is the same way a biological brain is organized, namely according to the same rules. Now, how do we implement these rules in a technical product that resembles a computer? Well, first we use the same materials and production equipment, namely semiconductors in foundries. Second, we have to put all the neurons into silicon, and this is where the magic word comes in: digital circuit design. How do we do that? Well, using VHDL or Verilog as a design method. But wait, those aren't programming languages, they're hardware description languages where "things" are processed in parallel and simultaneously, just as neurons act.
We also need to give each neuron the ability to make new connections to other neurons. And we need to give neurons the ability to organize themselves into groups and form networks, the way a self-organizing human society works in a village or small town.
This is where we're stuck, because we're not able to break away from the computer model. And there is another reason: we are forced to use the term "enumeration" as a basic type to describe complex objects.
As a result, we cannot use mathematics, which causes us much discomfort and confusion. We also cannot use the basic rules of mathematics, such as addition and subtraction, because the basic unit is not the same as the mathematical unit, which says that everything is the same. Red and blue are different, unlike numbers like 1 and 2. And we are confronted with real objects where measurement is of no help. Instead, we have to analyze, categorize, evaluate and decide in a fraction of a second, because we have a body and are confronted with different situations with many types of objects that we encounter in different landscapes.
And then comes the crucial argument for enabling AI based on neural networks: the properties of self-learning and self-organization of accumulated knowledge.
There are no blueprints for how to implement these kinds of capabilities, as this requires some assumptions and constraints about how knowledge is organized in the first place. The fields of science known today are not helpful in this matter, because the internal organization of the brain is completely different.
Only, the brain beats any computer in things like logic, analysis, decision, etc. There is only one thing to do: We need to change our minds about how the world and especially our brain is structured and how it works internally, give AI an appropriate body, and give it complete autonomy. The rest will be a self sustaining process.
This is exactly what our company is developing at the moment, a cognitive AI.