The full text 2629 words, when learning is expected to grow 8 Fenzhong
Figure source: Unsplash
The summer of 1956, a conference held in Dartmouth (Dartmouth) University, "Artificial Intelligence" (artificialintelligence) was first proposed, many years later that the meeting has also been identified as a starting point for the study of global artificial intelligence. The spring of 2016, a AlphaGo century war with the human world's top players on Li Shishi, the advent of a new wave of artificial intelligence.
Experienced two ups and downs, artificial intelligence began a new round of outbreaks. Now, with the integration of artificial intelligence into their enterprise systems, scientists will turn their attention to new and innovative artificial intelligence.
That meta-learning field. In simple terms, meta-learning is learning to learn. Humans have a unique ability, can learn in any situation or environment. People learn to adapt. People will think of ways to learn. Artificial intelligence in order to have the flexibility of this learning requires universal artificial intelligence.
In other words, AI need for an effective and efficient way to understand the learning process.
Artificial intelligence very different way of learning and human
Limited human and artificial intelligence learning process the most important differences.
Human capacity constraints. The human brain is limited, and time is limited, therefore, the ability of the human brain to adapt is limited. The human brain make full use of each information received, and then to develop the ability to train a large number of world model. Human beings are universal learners. If people's learning process efficient, so you can quickly learn all the disciplines. But not everyone is a quick learner.
In contrast, artificial intelligence have more resources, such as computing power. However, the study of artificial intelligence data than the data used in the human brain much more. These massive data processing requires enormous computing power.
Figure source: Unsplash
Meanwhile, with the task becomes more complex artificial intelligence, computing power is growing exponentially. Each AI inference made (across multiple data repositories) algorithms rely on establishing a connection between the different data segments. If the algorithm is not valid for a given set of data, then the exponential growth of computing power will show. Today, no matter how rotten Street computing power, exponential growth is not what we want solutions.
This is why, at present it will be used as artificial intelligence learner particular purpose. By learning from the similarity of the relevant data, artificial intelligence and can efficiently process the data from which to infer, without having to spend too much cost.
The importance of learning to learn Artificial Intelligence
When technicians tried to solve the exponential growth of computing power, there has been "learning" problem, because the artificial intelligence began to draw inferences from the increasingly complex data.
To prevent the exponential growth of computing power, artificial intelligence must identify the most effective learning path, and remember the path. Once the algorithm for different types of questions to determine the learning path, artificial intelligence can learn by choosing the path, follow the path of learning, adjusted according to changes in the learning path, so self-regulation, and dynamically guide them to find solutions.
This leads to the next problem of artificial intelligence: "multitasking."
As technicians began offering related but unordered task for artificial intelligence, "multitasking" came into being. If independent tasks can be executed at the same time how to do? If you perform certain tasks in artificial intelligence, knowledge and data to help it perform other tasks it?
"Multitasking" issue will be a problem, "learning to learn" is elevated to a new height.
In order to be "multi-task" operation, AI can be evaluated in parallel need independent data sets, but also associated data connection segment and infer the data. When artificial intelligence steps to perform a task, you need to constantly update their knowledge, so that you can use and apply this knowledge in other cases. Since tasks are interrelated, the assessment of the task will require to complete the entire network.
MultiModel is an example of Google's artificial intelligence system, the system learned while performing eight different tasks. The analog system of the brain to perceive information, can detect objects in an image, captions, speech recognition, translation between four pairs of the language selection and performs a syntax analysis. The system excels in multitasking joint training. Neural networks learn from the data in different fields.
In order to make it more adaptable artificial intelligence will need to learn multi-tasking. Artificial Intelligence as adaptive learners an application is in the field of robotics, namely robots replace humans learn to perform tasks in dangerous situations. For example, when monitoring or capture the situation changes, mechanical dogs will be able to adapt to circumstances, without the need to follow a specific command mankind.
Figure source: Unsplash
Artificial intelligence can learn how to become a universal learners do?
As we have seen from Google's MultiModel in that, of course, artificial intelligence can learn to become such a common human learners. However, it still needs some time to achieve. It consists of two parts: the RMB yuan reasoning and learning. Yuan reasoning focuses on the effective use of cognitive resources. Meta-learning focused on the effective use of limited human resources and limited cognitive data unique ability to learn.
In the Yuan reasoning, which is a key element of strategic thinking. If artificial intelligence inference can be drawn from the different types of data, whether it can adopt effective cognitive strategies do in different situations?
It is currently conducting research to identify gaps between human and artificial intelligence, cognitive learning, such as knowledge of the internal state of the memory of accuracy or confidence. However, in the final analysis, it depends on the reasoning yuan grasp the overall situation and strategic decisions. Strategic decisions comprises two parts: select from the available conventional strategies, see different strategies according to the situation. These are the areas of research dollars reasoning.
In the meta study, which is a key part of bridging the gap between the use of a large amount of training data and limited data model train model. Models must be adaptive in order to make accurate decisions based on little information across multiple tasks.
In this regard, there are different solutions. Some models through learning parameters of human learners to find a set of parameters that works properly in different tasks to achieve. Some models define the best learning space, such as metric spaces, learning may be most effective in this space. Some models, such as small sample meta-learning, learning their baby's learning algorithm to learn by imitating the least amount of data. These are the areas of research dollars to learn.
Yuan Yuan reasoning and artificial intelligence to learn just become part of GM learners. Put them with information from the motor and sensory processed together, the learner can more like human AI.
Figure source: Unsplash
AI still learning to become more like humans
Extensive research has become like a human learners need to learn a wide range of human and artificial intelligence methods on how to mimic human learning.
To adapt to new situations, such as having a "multitasking" capabilities and the ability to use limited resources to make a "strategic decision", which is a few obstacles to AI researchers during the study need to cross.
In human endeavor, artificial intelligence, learning ability is evolving, although compared with humans still a wide gap, but I believe the gap will continue with the progress of human technology and gradually reduce and eventually reached a shocking height, let us wait and see ~
Message thumbs attention
We share with AI learning and development of dry cargo
welcome attention of the whole class platform AI hanging from the media "read core technique"
(Add small series micro letter: dxsxbb, adding readers circle to discuss the freshest artificial intelligence technology together oh ~)
