Shock! “Next generation AI” is actually a football coach? ! [Manual dog head]
Editor’s note: This article is from WeChat public account “Brain Pole” (ID: unity007), author Tibetan fox. Authorized to reprint.
Remember the fear of being dominated by Alpha dogs? The latest “AI Horror Story” is that Google teaches AI to play football and create “AI version of Beckham.”
In fact, this football artificial intelligence research project, Google Research Football, appeared on the Google report as early as June this year. Later, on Github, the beta version of the football environment game was released in open source. Code.
With the Google “AI top stream” coffee table, plus today’s information efficiency, it is now suddenly a wave of popularity, the reflex arc is a bit too long. Is it that the media are collectively ignoring hot spots and changing into cold meals?
The answer may be in the imaginary flooding headline of “Next Generation AI,” which sounds sexy when used in an ancient Internet.
The AI that had surpassed the human e-sports team in “StarCraft” and “DOTA2” had not won this award. Playing football will lead the future. Is the difference between AI and AI so big?
How big is the upside for this generation of AI?
The public has become more and more aware that AI can’t be magical. However, I will revert my time back to 2016. I believe that most people will not deny that they are being replaced by AI. They are so scared by the robot network Red Sophia that they feel that “Terminator”, “The Matrix” and “Western World” will sooner or later. Come, bothIt is a journey of heart and mind.
The traditional “AI horror story” declared bankruptcy. On the one hand, it benefited from the accumulation of technology giants (怼), and also related to the technical limit based on deep learning.
For example, based on huge data calculations, training AI to play video games often costs hundreds of thousands of dollars; for example, black box, there is no deep learning structure (convolution, RNN, LSTM, GAN, etc.) You can explain your own decisions, and secretly engage in discrimination, swearing, and inventing new languages. In other words, you will only do “fill in the blanks” and face performance when you need to quote common sense, consensus, reasoning, etc. Like a mental retardation, it is easy to be fooled. For example, the printed face is recognized as true, or the IQ is not as expected. Medical diagnosis, robotics, automatic driving, etc. are always progressing slowly… Musk, who has always been anti-bone, has recently launched A new unmanned approach based on computer vision perception.
Overall, what deep learning can really do is to achieve a mapping between two spatial things given a large amount of artificially annotated data. AGI’s strong artificial intelligence, which people envisioned, is really “a half-hearted” and extremely far away.
So, the famous “singing AI” expert Filip Piekniewski claimed to give the “AI winter” pot to deep learning, although some sensational, it is not a point to point out a practical and serious problem – if you study deeply The basic AI application will not continue to improve, so the relevant industries have come to the end of the road (especially those to VC projects), it is also a matter of time soon~
DL is not the ultimate algorithm, so keep on indulging in games
Since this is the case, how do you do AI? In theory, there are two angles: one is the self-evolution of deep learning, introducing new technologies on the basis of the original to make up for some congenital deficiencies; the other is looking for “prepared tires”Support other genres in the AI field.
At present, technology companies are indeed extremely eager for the emergence of variables, but they prefer to be a moderate “improvementist”. After all, “completely overthrowing the corrupt regime” requires a long process of training successors.
Taking Google Football Engine as an example, let the agent use the reward mechanism to get dynamic strategies to learn rules and kicking skills (enhanced learning).
However, it is a bit of a nucleus to call it “the next generation of AI.”
First of all, “Gameplay Football” is not completely free from the shackles of deep learning. Based on the strength of the opponent, the system proposes a benchmark problem of three versions: simple, ultimate, and difficult. The simple level of competition applies a single machine algorithm, while the difficulty level is handled by a distributed deep learning algorithm.
Moreover, the training method used by the system (ie, reinforcement learning), and OpenAI Five defeated the world-class eSports team OG in the game Dota 2, and the training used by deepmind in the Warcraft Man-machine battle. There is no essential difference between the ways in which the agent learns to interact with the environment and solve complex tasks in complex real-time strategy games.
At the same time, as a branch of machine learning, reinforcement learning is still far away from AGI. Yann LeCun and Hinton of the Big Learning Big Three believe that the current technology used to achieve “artificial intelligence effects” does not work for real (real) artificial intelligence. Just like how to optimize the core technology of the carriage, it can’t make a car.
What’s more, there are many similar machine learning methods that make up for the lack of deep learning.
Small sample learning, unsupervised learning, free from the need for large-scale data sets and human expert supervision, improve the efficiency of independent training; meta-learning solves the intelligent skills of deep learning trainingA single, lack of common sense. In 2015, deep learning Hinton also proposed a black technology, Knowledge Distillation, to enhance deep learning on large-scale computing clusters by migrating knowledge and using trained large models to obtain small models more suitable for reasoning. Training performance.
All in all, the so-called “next generation AI”, the core is to make up for the lack of deep learning in understanding, multi-modal bionics, application cost performance. As a transitional solution, this “deep learning +” estimate will continue for a long time. However, the expectations of real AGI are still far from each other.
Following the next generation of AI, perhaps going to a broader technology area
Most of the AI product ideas we saw today were built using the DL (Deep Learning) + GOFAI (Good Old Fashioned AI) model. That is to say, deep learning is combined with other algorithms to make “AI” go all the way.
But many scientists are thorough “revolutionaries” and think of a lot of new ways to help AI, which may also hide the feasibility of breaking.
For example, Hinton tried to subvert the traditional deep learning algorithm through the capsule network Capsule Networks, replacing the single neuron node of the traditional neural network with the neuron vector, and let the different neurons carry the information of different attributes to the next. A layer of computing has proven to be able to automatically extend the knowledge learned to different new scenarios, just like the human visual system, which is considered to be the key technology for AI to be given common sense reasoning in the future.
Other experts insist that logic-based symbology can achieve AI reasoning, and some scholars and startups are using Prolog, a semiotic-based programming language, to develop new tools. In theory, you can train with very little data, handle facts and concepts yourself, and then automatically generate factual inferences.
But overall, the AI genre of other branches wants to incite the mainstream status of Deep Learning 2.0, which is still difficult. In addition to the industry’s aggressive investment in deep learning and derivative technology, the US Defense Advanced Research Projects Agency DARPA even prepared a “machine knowledge”The (Machine Common Sense) program is designed to advance and share technical ideas that mimic human common sense reasoning, with a total investment estimated to be approximately $60 million.
As a benchmark for deep learning and extension technology, its commercial potential, even if it is “backing the mountain”, there are several years of good prospects. However, it must be acknowledged that in the face of its own bottleneck, the public’s adrenaline and technical expectations are beginning to return to normal, and even a little aesthetic fatigue. Technical experts will not make a big event anymore, and the tech editors who love the “AI Ghost Story” will be forced to bald…
Where the AI direction of the next-generation industry worth exploring is going, I am afraid it is far from our existing knowledge. After all, every huge change in the world always begins in some corners of the neglected technology. In addition to continuing to challenge the dome of technology, Google seems to have no choice.