Lecun published a research plan for the next decade: AI autonomous intelligence with a 62 page paper

Abundant color From the Aofei temple
qubits | official account QbitAI

During this period , About “AI Where to go in the future ” Discuss , It can be said that it is becoming more and more intense .

First, Meta Be exposed AI Major restructuring of relevant departments , There's Google again AI A big discussion on whether you have personality , Almost every discussion can see Yann LeCun The figure of .

Now? ,LeCun Finally, I can't sit down .

He wrote a long article 62 The latest paper on page , He introduced in detail what he would do in the next ten years AI Research ：

Autonomous machine intelligence （Autonomous Machine Intelligence）.

LeCun Express , Most practitioners will not publish their research contents in advance “ Academic atmosphere ” Next , His action is very special .

The reason is , In addition to carrying forward the spirit of open scientific research , It is also to call on more people to join in , Study together .

that , What he said about autonomous artificial intelligence , What is it , And how to carry out ？

Can simulate the operation of the world AI

In the paper ,LeCun First, I gave an example ：

A young man can be the fastest in 20 Learn to drive within hours ;

One of the best auto drive system systems in the world today , It takes millions or even billions of labeled training data , And millions of times of intensive learning in the virtual environment —— It is not at all human level .

From this example, we can draw , Although we have made a lot of progress in artificial intelligence , but To create a world that can really think and learn like human beings AI Not even close .

LeCun The proposed autonomous artificial intelligence is to solve this problem .

In his opinion , Yes “ The world model ”（ An internal model of how the world works ） The ability to learn can be key .

as everyone knows , Humans and other animals can always interact with a small amount through observation , You can learn in an unsupervised way A lot of Background on how everything works in the world .

This knowledge is what we call common sense , Common sense is what constitutes “ The world model ” The basis of .

With common sense , We can also act in unfamiliar situations . For example, the young man who had never driven a car at the beginning , Hit the snow , You don't have to teach me how to drive slowly on such a slippery road .

Besides , Common sense can also help us fill in the lack of information in time and space . For example, a driver hears the collision sound of metal and other substances , Even if you don't see the scene , You can also know that there may be an accident .

On top of these concepts ,LeCun The first challenge of building autonomous artificial intelligence ：

How to design a learning paradigm and architecture , Let the machine learn by self-monitoring （ That is, you do not need to label data ） Way to learn “ The world model ”, Then use this model to predict 、 Reasoning and action .

ad locum , He recombined cognitive science 、 Systems neuroscience 、 optimum control 、 Strengthen learning and “ Tradition ” Ideas proposed in various disciplines such as artificial intelligence , And combine them with new concepts in machine learning , Came up with a An autonomous intelligent architecture composed of six independent modules .

among , Each module is differentiable , Each can easily calculate the gradient estimation of an objective function relative to its own input , And propagate the gradient information to the upstream module .

Six module autonomous intelligence architecture

LeCun The six modules envisaged are ：

1、 Configuration module ： Responsible for implementing control . Given the task to be performed , It can adjust the parameters of other modules , Preconfigured awareness modules for tasks 、 The values of the other three modules, such as the world module .

2、 Perception module ： Responsible for receiving signals from sensors and estimating the current state of the world .

3、 World model module ： Is the most complex part of this architecture . It does two things ：

（1） Estimate the missing information about the state of the world that the perception module cannot provide ;

（2） Predict possible future states . Because the world is full of uncertainty , The module must be able to cover a variety of possible predictions .

4、 Cost module ： Used to calculate scalars （scalar） Output , It can predict the discomfort degree of the agent （discomfort of the agent, Damage to the agent 、 Violation of hard coded behavior constraints, etc ）.

This module has two sub modules ：

（1） Internal cost module （cost）, For instant calculation “ Discomfort ”;

（2） Judge (critic)： Predict the future value of the intrinsic cost module .

5、 Action module ： Used to calculate the action sequence to be implemented . The action module can find an optimal action sequence that minimizes the future cost module , And in a way similar to classical optimal control , Output the first action in the optimal sequence .

6、 Short term memory module ： Track current and projected world conditions and associated costs .

among , For the core of this architecture —— World module , The key challenge is how to make it possible to express multiple reasonable predictions .

Besides , It is learning the abstract representation of the world , And learn to ignore irrelevant information , Keep only the most useful details .

Like driving , Just predict what the cars around the driver will do , It is not necessary to predict the detailed position of each leaf in the trees on both sides of the road .

Regarding this ,LeCun And a possible solution ：

Joint embedded prediction architecture (JEPA), Use it to deal with the uncertainty in the forecast .

meanwhile , He also proposed using non comparative self supervised learning to JEPA Training , And the classification of prediction from different time scales JEPA, It can disassemble complex tasks into a series of less abstract subtasks .

AI There are still many problems to be solved

LeCun Express , For decades to come , Training such a world model is the biggest challenge that artificial intelligence must face in order to make a breakthrough .

For now , To implement the above Architecture , There are still many aspects to be defined ： For example, how to train accurately critic、 How to construct and train configurators 、 And how to use short-term memory to track the state of the world , And store the state of the world 、 The history of actions and associated intrinsic costs critic……

besides ,LeCun It is also pointed out in the paper that , For the future research of autonomous artificial intelligence ：

（1） It is necessary to scale up the model , But not enough ;

（2） The reward mechanism is not enough , Self supervised learning based on observation is a more effective way ;

  (3) Reasoning （reason） And planning （plan） In essence, it all comes down to inference （inference）： Find a series of actions and potential variables , To minimize （ It's very small ） The goal is . This is also a way to make reasoning compatible with gradient based learning .

 （4） In this case , There may be no need for an explicit symbol manipulation mechanism .

More details can be found in the original paper ：
https://openreview.net/forum?id=BZ5a1r-kVsf

Reference link ：
[1]https://twitter.com/ylecun/status/1541492391982555138
[2]https://ai.facebook.com/blog/yann-lecun-advances-in-ai-research/

—  End  —

「 qubits · viewpoint 」 Live registration

Why? “GPU+” It is more suitable for high performance and versatility AI Computational conundrum ？ The public attention “GPU+” Where is the core competitiveness of architecture ？ Use “GPU+” How powerful can the architecture chip be ？

6 month 30 On Thursday , Participate in the live broadcast , Answer for you ～

Focus on me here , Remember to mark the star ～