The Opportunities and Challenges of AI Public Chain
AI’s perception and sensing of information, information conversion and transmission, and information processing are the expertise of artificial intelligence. For the artificial intelligence blockchain, the perception and sensing layers are spontaneously distributed, while the other two stages have not been implemented at current, which is also the most root of the problems of scalability, privacy and scalability. Looking forward to the future of the artificial intelligence blockchain, we hope that it can become the spine and nervous system of the blockchain, accurately and effectively deal with the three major problems of the blockchain which mentioned above.
1 Opportunities
Benefit from the unique attributes of artificial intelligence by introducing AI technology into the blockchain.
First, it has a self-learning function. For example, when image recognition is implemented, only a plurality of different image templates and corresponding recognition results are input into the artificial neural network, and the network will learn to recognize similar images through the self-learning function. The self-learning function is especially important for prediction. It is expected that the future artificial neural network computer will provide economic forecast, market forecast and benefit forecast for human beings, and its application prospect is very ambitious.
Second, it has the associative memory function which can be achieved using a feedback network of artificial neural networks.
Third, it has the ability to find optimized solutions at high speed.
Form 1 summarizes the corresponding relationship between the various attributes of AI and the improvement of various aspects of the blockchain.
AI’s perception and sensing of information, information conversion and transmission, and information processing are the expertise of artificial intelligence. For the artificial intelligence blockchain, the perception and sensing layers are spontaneously distributed, while the other two stages have not been implemented at current, which is also the most root of the problems of scalability, privacy and scalability. Looking forward to the future of the artificial intelligence blockchain, we hope that it can become the spine and nervous system of the blockchain, accurately and effectively deal with the three major problems of the blockchain which mentioned above.
1) Learning Function
Reinforcement learning is characterized by exploratory interaction with the environment to determine and optimize the choice of actions to achieve so-called sequence decision tasks. In this task, the learning mechanism causes changes in the state of the system by selecting and executing actions, and it is possible to obtain some kind of reinforcement signal (immediate return) to achieve interaction with the environment. reinforcement signal is a kind of reward and punishment for the system behavior. The goal of system learning is to select a suitable action strategy, namely selected the most suitable action method in different state, so that the generated action sequence can obtain the optimal result (such as the cumulative immediate return is maximum).
In the comprehensive classification, experience-inductive learning, genetic algorithm, joint learning and reinforcement learning are all inductive learning, in which experience-inductive learning adopts symbolic representation, while genetic algorithm, joint learning and reinforcement learning adopt sub-symbol representation and analytical learning belongs to deductive learning.
In fact, the analogy strategy can be seen as a synthesis of inductive and deductive strategies. Therefore, the most basic learning strategies are only inductive and deductive.
From the perspective of learning content, it using inductive strategy which is based on the induction of input, the knowledge learned is obviously beyond the scope of the original system knowledge base, and the learned results change the knowledge deduction closure of the system, so this type of learning can be called knowledge level learning. While learning with deductive strategies can improve the efficiency of the system, but still be implied by the knowledge base of the original system, that is, the learned knowledge fails to change the system’s deductive closure, so this type of learning is called symbol level learning.
2) Associative Memory
Because the memory data has been solidified into a neural network structure with a certain function, this whole trained neural network is memory. Human brain memory cannot leave the neural network alone. It is necessary to reconstruct the neural network to transplant memory. It is not as simple as downloading a computer hard disk. It is functional. People can easily remember something and forget it for one time, which is different from the artificial neural network training. But how the neurons and connections are remembered, the mechanism is still determined by the Human Brain Project. It is certainly possible to record a simple message with the weights of multiple neurons. This is also a coding expression, while it is consistent with the simple information stored in the memory. The memory stores a simple message, it can be multiple bytes and many bits. You can treat each byte as a weight, multiple bytes as multiple weights, and part of the memory can be seen as a neural weight network without active response. Or further, part of the binary bit is regarded as a unit of neuron weights with only active and suppressed states. Multiple bits can encode and represent complex information, which can also explain some biological neural network research. Therefore, the memory can approximately correspond to the memory of the biological neural network under certain circumstances.
Ability to find optimized solutions at high speed.
Finding an optimal solution to a complex problem often requires a large amount of computation. Using a feedback artificial neural network designed for a problem, the computer’s high-speed computing power can be used to find an optimal solution.
3) Programmability
Artificial intelligence has basic programmability, embedding artificial intelligence programs into the user’s wallet, and updating the artificial intelligence in the wallet every time the user synchronizes the nodes, so that the things learned by artificial intelligence are synchronized with each other through continuous optimize and strengthen the control of the overall smart contract. Make artificial intelligence grow up quickly, replace people to do things that people don’t want to do and things that people can’t do.
2 Challenges
The opportunities presented by artificial intelligence do not mean that several algorithms can be well integrated into the blockchain. In fact, there are many challenges, and none of the existing artificial intelligence can be applied to the blockchain to solve the above.
Let the computer learn without the help of a human teacher.
The most successful machine learning method to date has been called supervised learning, which is very similar that teacher pointing at something and telling us that the names. Every time you learn a new task, the system basically has to start from scratch, requiring humans to participate for a long time to a large extent.
1) Theoretical Challenges
At present, the deep neural network is modeled by mimicking the reticular neural structure of the human cerebral cortex. The actual constructed models are simplified MNN, and the nonlinear mapping relationship is mainly expressed by the connection between adjacent layers. If a connection is also established between non-contiguous layers or peer neurons, can the learning and presentation ability of the deep network be improved? Can we find a basis from neurology? Can we construct a deep neural network to effectively deal with machine learning problems equivalent to human intelligence? How to construct a deep neural network, so that the physical meaning of each layer extraction feature can be clearer? Relative to the mainstream can a two-stage training algorithm find a completely unsupervised online training algorithm?
2) Modeling Challenge
If a non-contiguous layer or a peer neuron is allowed to be connected, how should the deep neural network model be constructed? How can the deep model be improved so that the input data can be input into the model with simple preprocessing and can directly process multimodal data? How to construct a deep model to reduce its dependence on tagged data? How to transform a deep model to achieve parallel acceleration?
3) Engineering Realization Challenges
The deep neural network training time is too long, and it is easy to over-fit, which weakens the modeling and promotion ability. It is an important issue needs to be solved that how to transform the deep neural network training algorithm to make a good model promotion performance which can quickly converge to the optimal solution and reducing the training time. How to transform the deep model to apply to multiple types of input data or even multi-modal mixed data? How to transform the deep model to effectively combine GPUs and parallel acceleration technologies such as distributed computing?
3 Related Exploration
We propose gcForest (multi-Grained Cascade forest) and a new Decision Tree ensemble method. This method generates a deep forest ensemble method and use gcForest make feature learning with cascade structure. When the input has a high dimension, its multi-grain scanning can further enhance its feature learning ability, which is expected to make gcForest notice contextual or structural aware. The number of cascades can be adjusted as appropriate, so that gcForest also exhibits excellent performance with only small data. It should be noted that gcForest’s hyperparameter are much less than deep neural networks, and gcForest is quite robust to hyperparameter, so in most cases, even if you encounter different data in different fields, you can use the default settings to get good results.
MAI Telegram (English): https://t.me/MaiEN8888
MAI Telegram (Chinese): https://t.me/MAI888888
MAI Twitter: https://twitter.com/MaiPublicChain
MAI Reddit: https://new.reddit.com/user/MAI_AI
MAI FACEBOOK: https://www.facebook.com/MaiPublicChain/
