Home  >  Article  >  Technology peripherals  >  As long as the model is large enough and the samples are large enough, AI can become smarter!

As long as the model is large enough and the samples are large enough, AI can become smarter!

王林
王林forward
2023-04-29 15:25:061468browse

​There is no difference in mathematical mechanism between the AI ​​model and the human brain.

As long as the model is large enough and the samples are large enough, AI can become smarter!

The emergence of chatGPT has actually proved this point.

1. The underlying details of AI and the human brain are based on if else statements

logical operations, which are the basic operations that generate intelligence.

The basic logic of the programming language is if else, which divides the code into two branches based on conditional expressions.

On this basis, programmers can write very complex codes and implement various business logics.

The basic logic of the human brain is also if else. The two words if else come from English, and the corresponding Chinese vocabulary is if...else...

When the human brain thinks about problems It’s also such a logical idea, and it’s no different from a computer in this regard.

As long as the model is large enough and the samples are large enough, AI can become smarter!

if else statement, the core of logic

The "if else statement" of the AI ​​model is the activation function!

A computing node of the AI ​​model, we can also call it a "neuron".

It has an input vector X, a weight matrix W, a bias vector b, and an activation function.

The activation function is actually an if else statement, and the linear operation WX b is a conditional expression.

After activation, the code of the AI ​​model is equivalent to running in the if branch, and when not activated, it is equivalent to running in the else branch.

The different activation states of multi-layer neural networks are actually binary encodings of sample information.

As long as the model is large enough and the samples are large enough, AI can become smarter!

Deep learning is also a binary encoding of sample information

The encoding of sample information by the AI ​​model is dynamic and parallel, not the same as the CPU code They are static and serial, but their underlying basis is if else.

It is not difficult to implement if else at the circuit level. It can be implemented with a triode.

2. The human brain is smarter than computers because humans obtain more information.

The human brain is acquiring information from the outside world all the time and updating its own information all the time. "Sample database", but the program code cannot update itself, which is why many people can do it but computers cannot.

The code of the human brain is alive, but the code of the computer is dead.

Of course "dead code" cannot be smarter than "live code", because "live code" can actively find the bugs of "dead code".

According to the continuity of real numbers, as long as the information encoded by "dead code" is countable, then it will always have bug points that cannot be encoded.

This can be mathematically supported by Cantor's three-point set.

No matter how many ternary decimal digits we use to encode real numbers in the interval [0, 1], there is always at least one point that cannot be encoded.

So when two people argue, they can always find points to argue with.

But once the computer code is written, it cannot be updated automatically, so programmers can come up with various ideas. A way to trick the CPU.

For example, Intel's CPU originally required switching task gates when switching processes, but Linux came up with a way to only switch the page directory and RSP register

In the view of Intel CPU , the Linux system has been running the same process, but it is not. This is called process soft switching.

So, as long as the circuit of the CPU is fixed, the information encoded by the CPU will also be fixed.

The information encoded by the CPU is fixed, so the information it cannot encode is unlimited and can be used by programmers.

The reason why programmers can use this information is because the programmer's brain is alive and can dynamically update samples.

3. The emergence of neural network has changed this situation.

Neural network is really a great invention. It realizes dynamic information update on fixed circuits.

The information that all written programs can process is fixed, including CPU circuits and codes of various systems.

But this is not the case with neural networks. Although its code is written, it only needs to update the weight data to change the logical context of the model.

In fact, as long as new samples are continuously input, the AI ​​model can continuously update the weight data using the BP algorithm (gradient descent algorithm) to adapt to new business scenarios.

Updating the AI ​​model does not require modifying the code, but only requires modifying the data, so the same CNN model can recognize different objects if it is trained with different samples.

In this process, both the code of the tensorflow framework and the network structure of the AI ​​model remain unchanged. What changes is the weight data of each node.

Theoretically, as long as an AI model can crawl data through the network, it can become smarter.

Is this fundamentally different from people watching things through a browser (thus becoming smarter)? It seems not.

4, as long as the model is large enough and the samples are large enough, maybe ChatGPT can really challenge the human brain

The human brain has 15 billion neurons, and human eyes and ears are constantly changing Of course, the AI ​​model can also do this by providing it with new data.

Perhaps compared to AI, the advantage of humans is that the "industrial chain" is shorter

The birth of a baby only requires its parents, but the birth of an AI model obviously does not require one or two programs Members can do it.

There are more than tens of thousands of people manufacturing GPUs alone.

The CUDA program on the GPU is not difficult to write, but the industry chain of GPU manufacturing is too long, far inferior to the birth and growth of human beings.

This may be the real disadvantage of AI compared to humans.

The above is the detailed content of As long as the model is large enough and the samples are large enough, AI can become smarter!. For more information, please follow other related articles on the PHP Chinese website!

Statement:
This article is reproduced at:51cto.com. If there is any infringement, please contact admin@php.cn delete