Lecun, a Turing Award winner, pointed out that the future of AI lies in self-learning, and the company has embarked on the journey

  In recent days, , Turing award winner Yann LeCun Published for 62 Page paper , Summarized his recent 10 For years AI Thinking about the general direction of industry development , And for the future 10 It was proposed that “ Autonomous Agents ” framework （autonomous intelligence） New directions for . He thinks that , Let the machine know more about the key to general intelligence , It's about “ independent ” Two words .

The dilemma of general artificial intelligence ： One cannot turn three

“ Universal Artificial Intelligence ”（AGI, That is, strong artificial intelligence ） The emergence of is still far away . The reason is , Lies in the present AI It still relies on massive training data , With tens of millions or even hundreds of millions of levels of passive supervised learning and training , To realize the reuse of a little bit of human cognitive ability .

At present AI It can use powerful computing power to process hundreds of billions of data that human beings cannot take into account , But in active comprehension , May even 2~3 Children are not as good as .

Like a teacher teaching , Supervised training and learning cannot exhaust all possibilities in a short time , And when the corresponding rules become complicated , Lack of corresponding data , The types of training data generated cannot be exhaustive ,AI You can't learn actively and effectively .

for example , The average young person 20 You can learn to drive in hours , Can learn to drive from various simple rules , Don't enumerate all the cases .

But with extremely complex road conditions and the possibility of accidents , An excellent automatic driving technology needs millions or even billions of labeled data , In the virtual environment, millions of reinforcement learning can be generated , And its reliability is not necessarily comparable to human .

How to make the machine have certain autonomy , From passively receiving data and training to active learning , Improve learning efficiency , This is the winner of the Turing prize Yann LeCun Consider the direction .

" Autonomous Agents " And autonomous learning ：AI The future direction

Yann LeCun Put forward “ Autonomous Agents ” The new architecture of , It contains six differentiable modules , Each module can easily calculate some objective functions , And the corresponding gradient estimation , And propagate the gradient information to the upstream module .

Which includes ：“ Configuration module ”、“ Perception module ”、“ World model module ”、“ Cost module ”、“ Action module " And " Short term memory module ”.

Configurator （Configurator） It is the top-level module in the new structure he conceived , Responsible for obtaining input from other modules , And adjust the parameters of other modules according to the task requirements , So as to pre configure perception （perception）、 The world model （world model）、 cost （cost） And participants （actor） The corresponding value of . It also makes the configurator module more like a brain center , Ability to output results more autonomously .

The concept of configurator module and meta learning in machine learning （Meta-learning） Consistent .Meta-learning Designed to make the machine learning to learn, That is to say, let the machine learn how to learn .

Unlike traditional machine learning, advanced pedestrians work for tuning participation , Then directly train the depth model under specific tasks ;Meta-learning It will let the machine learn all the parameters and variables that need to be set and defined manually in advance , This includes letting the machine learn how to preprocess data in advance 、 Select the network structure 、 Set super parameters 、 Define the loss function and so on , The experience gained from these learning histories gives machines meta knowledge , Thus only a very small number of samples are required in the future , You can quickly adapt to and master new tasks .

Meta-learning The idea of making the machine learning process more autonomous , Thus, it can well replace many jobs of human engineers .

LeCun Also put forward ,AI Want to break through the current bottleneck , Machines must be allowed to learn autonomously , So that we can fill in the missing information , Predict what will happen , And predict the impact of action .

Deal with the future of large-scale refinement with autonomous learning

meanwhile , Due to the emergence of large-scale and large-scale application scenarios , And the demand for refined personalized algorithms surges , bring AI The production of models shows a large and diversified trend , In different industries 、 Different scenes 、 Different environments 、 Run on different devices , Algorithms need large-scale personalized development .AI The number and scale of , Expect explosive growth in the future .

Take the visual algorithm as an example , Previous algorithms were only deployed to a limited number of traffic or property surveillance cameras , Only undertake face recognition 、 Vehicle and other general functions . But in recent years , More customized 、 Refined algorithm requirements , Has begun to put on the table of various wisdom scenes . In the future, with the increase and popularization of hardware shipments ,AI It will penetrate into the details of various industries , such as ：

Wood processing plants in Guangdong will use AI, Used to identify the accumulation of wood debris , Timely remind workers to clean up and avoid potential fire hazards ; Dairy farms in Inner Mongolia will use AI, To identify the accumulation of cow forage and water the cow's skin regularly ; Workshop factories in Shanghai need to use AI, Help the administrator judge whether the employee is over tired ; Even your neighborhood will use AI, When the old man falls 、 When garbage overflows ,AI Can respond at the first time .

All these very large-scale deployment scenarios require , bring AI Can no longer be like water 、 electric 、 Standardized supply like coal ,AI Algorithms will depend on each industry 、 Each scene 、 Every need 、 Every design and iteration is the same , Become a mass of fragmented and refined customized products .

Under the trend of super large-scale refinement ,Yann LeCun It is also believed that the use of general large models 、 By pure data stacking 、 Let the machine learn passively 、 The old way of repeatedly manually adjusting parameters , Has been unable to get through , He also pointed out that making machine learning more autonomous , It will be the future AI The most important direction of development .

Reach the ground together AutoML Automated training

Give Way AI Learn to “ Autonomous Learning ” In this direction , Gongda is committed to using AutoML technology , take LeCun And meta-learning Put our ideas into practice .

In human intensive and highly repetitive AI Under the R & D process , Gongda relies on industry leaders AutoML technology , use AI Training AI, Get rid of the common ridicule in the industry as “ Only man , No intelligence ” The tedious research and development AI technological process , With unprecedented acceleration kinetic energy , by AI Industry brings qualitative changes .

Data engineering through semi supervised learning ; Automatically adjust the model structure 、 Hyperparametric 、Loss Function and Head Methods such as ; And automatically adjust the neural network structure according to the chip operator , The original training time can be changed from 3 to 6 Months were reduced to two weeks , One click to send to the cloud / Side multi terminal equipment . And no algorithm team is needed 0 Code experience can also be used , Just upload the data and click on the training to generate a AI Algorithm .

After deployment , There is no need for professional Algorithm Engineers to participate in the operation and maintenance team , Sustainable and automatic access to field data , Iteratively optimize the algorithm , Make the accuracy close to the ideal value of the scene .

At present, we have reached a total of AI Widely used in smart cities 、 Smart industry 、 Wisdom park 、 Smart buildings and other scenes .

Reference resources ：A Path Towards Autonomous Machine Intelligence | OpenReview