News classification
Contact us
- Add: No. 9, North Fourth Ring Road, Haidian District, Beijing. It mainly includes face recognition, living detection, ID card recognition, bank card recognition, business card recognition, license plate recognition, OCR recognition, and intelligent recognition technology.
- Tel: 13146317170 廖经理
- Fax:
- Email: 398017534@qq.com
Artificial Intelligence Research and Application
Artificial Intelligence Research and Application
Recently, there are many controversies about the definition of artificial intelligence. Some people think that artificial intelligence is "cognitive computing" or "machine intelligence", while others put it and the "machine learning" concept confused. However, artificial intelligence is not specific to a technology, it is actually a wide range of disciplines composed of a number of areas, including robotics and machine learning. The ultimate goal of artificial intelligence is to replace the human machine to complete the task requires cognitive ability. In order to achieve this goal, the machine must automatically learn to master the ability, not just the implementation of the programmer to write the command.
Artificial intelligence has made amazing progress over the past decade, such as autopilot, speech recognition, and speech synthesis. In this context, the topic of artificial intelligence increasingly appears in the chat between colleagues and family, artificial intelligence technology has penetrated into the corner of their lives. At the same time, the popular media are reporting daily on artificial intelligence and technology giants, introducing their long-term strategy in the field of artificial intelligence. Some investors and entrepreneurs are eager to learn how to tap the value of this new field, most people still racked their brains to think about what the artificial intelligence will change. In addition, governments are working to address the social impact of automation (such as President Obama's departure speech).
Among them, the six areas of artificial intelligence in the future of digital products and digital services may have an important impact. The authors list these six directions, explain their importance, the current application scenarios, and list the companies and research institutions that are in use.
Intensive learning
Reinforcement learning is a way of learning through experimentation and error, inspired by the process of learning new skills. In a typical reinforcement learning case, the agent takes action by observing the current state and maximizes the outcome of the long-term reward. Each time an action is performed, the agent receives feedback from the environment, so it can determine whether the effect of the action is positive or negative. In this process, the agent needs to balance the experience to find the best strategy and explore new strategies both in order to achieve the ultimate goal.
Google's DeepMind team uses reinforcement-learning techniques in Atari games and Go games. In the real world, reinforcement learning has been used to improve the energy efficiency of Google's data centers. Reinforced learning technology for the cooling system saves about 40% of energy consumption. Reinforcement learning has a very important advantage, its agents can generate a large number of training data at low cost. Compared to the supervised depth learning task, this advantage is very obvious, saving a large amount of manual labeling data costs.
Applications: including automatic driving of urban roads; navigation of three-dimensional environment; multiple agents interact and learn in the same environment, etc.
Generate the model
Unlike the discriminant model used to perform the classification and regression tasks, the generation model learns a probability distribution from the training samples. By sampling from the high-dimensional distribution, the model is generated to output new samples that are similar to the training samples. This also means that if the training data of the generated model is the set of images of the face, the model obtained after the training can also output a composite image similar to the face. Details can be found in the article by Ian Goodfellow. The structure of GAN is very hot in academe because it provides a new idea for unsupervised learning. GAN structure uses two neural networks: one is the generator, which is responsible for random input noise data into new content (such as synthetic images), the other is the discriminator, responsible for learning the real picture and determine the generator-generated Whether the content is real oneself. Confrontation training can be considered a type of game, the generator must be repeated learning random noise data with meaningful content, until the judge can not distinguish the authenticity of the synthetic content. The framework is being extended to many data patterns and tasks.
Application: simulation of time series features (eg, planning tasks in reinforcement learning); super-resolution images; restoration of three-dimensional structures from two-dimensional images; generalization of small-scale annotation datasets; prediction of the next frame of video; generation of natural language Of the dialogue content; art style migration; voice and music synthesis
Memory network
In order for artificial intelligence systems to be able to adapt to a wide variety of environments like humans, they must continually acquire new skills and remember how to apply them in future scenarios. Traditional neural network is difficult to master a series of learning tasks. This shortcoming is what scientists call catastrophic forgetting. The difficulty is that when a neural network training for A task, if re-training it to solve the B task, then the weight value of the network model is no longer applicable to the task A.
At present, there are some network structure can make the model have different levels of memory capacity. Including a short-term memory network (a recurrent neural network) that can process and predict time series; the DeepMind team of micro-neural computers, which combines neural networks and memory systems to facilitate learning from complex data structures; progressive neural networks , Which learns the lateral associations between individual models and extracts useful features from these existing network models to accomplish new tasks.
Applications: training agents able to adapt to new environments; robotic arm control tasks; autopilot vehicles; time series prediction (eg financial markets, video prediction); understanding of natural language and forecasting below.
Microdata learning micro model
All along, the depth learning model needs to accumulate a lot of training data in order to achieve the best results. For example, a team that only participated in the ImageNet Challenge used 1.2 million manually tagged image training models distributed across 1,000 categories. Leaving large-scale training data, depth learning model will not converge to the optimal value, can not be in voice recognition, machine translation and other complex tasks to achieve good results. The growth of data demand often occurs when a single neural network model is used to process the end-to-end case, such as inputting the original speech segment, and outputting the converted text content. This process works with multiple networks to handle one intermediate result (for example, raw voice input → phoneme → word → text output). If we want to use the artificial intelligence system to solve the task of training data scarcity, I hope the training model used in the sample as little as possible. When the training data set is small, over-fitting, outliers interference, training set and test set inconsistencies and other issues will follow. Another approach is to migrate models that have been trained on other tasks into new tasks. This approach is called migration learning.
A related problem is the use of fewer model parameters to build a smaller deep learning architecture, while the model's effect is to maintain the best. The advantage of this technology is more efficient distributed training process, because the training process needs to reduce the transmission parameters, and can easily be deployed in the model memory size limited embedded hardware.
Applications: Training shallow models to simulate deep network models trained on large-scale labeled training datasets; building model structures that have comparable but lesser parameters (eg, SqueezeNet); machine translation
Learning / reasoning hardware
One of the catalysts for the development of artificial intelligence is the evolution of the graphics processor (GPU), which, unlike the sequential execution mode of the CPU, supports large-scale parallel architectures that can handle multiple tasks simultaneously. Since the neural network must be trained with large-scale (and high-dimensional) datasets, the GPU is much more efficient than the CPU. This is why since the first GPU training in 2012, the neural network model - after the release of AlexNet, GPU has become a veritable gold-plated spade. NVIDIA continues to lead the industry in 2017, ahead of Intel, Qualcomm, AMD and the rising star of Google.
However, the GPU is not designed for model training or prediction, it was originally used for video game image rendering. GPU with high-precision computing capabilities, but suffered memory bandwidth and data throughput problems. This opens up new areas for large companies such as Google and many small start-ups that design and manufacture processing chips for high-dimensional machine learning tasks. The improvement of chip design includes more memory bandwidth. Graph calculation replaces vector computation (GPU) and vector computation (CPU), higher computation density, lower energy consumption. These improvements are exciting because they ultimately feed back to the user: faster and more efficient model training → a better user experience → more users to use the product → gather larger data sets → improve through the optimization model Product performance. As a result, systems that train and deploy models faster take a significant advantage.
Applications: Rapid training of models; Low-power predictive computing; Persistent monitoring of Internet of Things devices; Cloud service architecture; Autopilot vehicles; Robots
Simulation environment
As mentioned earlier, it is challenging to prepare the training data for the artificial intelligence system. Moreover, to apply an artificial intelligence system to real life, it must have applicability. Therefore, developing a digital environment to simulate the real physical world and behavior will provide us with the opportunity to test the adaptability of artificial intelligence systems. These environments present the original pixel to the artificial intelligence system, and then take some action based on the set goals. Training in these simulated environments can help us understand the learning principles of artificial intelligence systems, how to improve the system, but also provide us with the model can be applied to the real environment.
Applications: Simulation Driving; Industrial Design; Game Development; Smart City
Artificial intelligence has made amazing progress over the past decade, such as autopilot, speech recognition, and speech synthesis. In this context, the topic of artificial intelligence increasingly appears in the chat between colleagues and family, artificial intelligence technology has penetrated into the corner of their lives. At the same time, the popular media are reporting daily on artificial intelligence and technology giants, introducing their long-term strategy in the field of artificial intelligence. Some investors and entrepreneurs are eager to learn how to tap the value of this new field, most people still racked their brains to think about what the artificial intelligence will change. In addition, governments are working to address the social impact of automation (such as President Obama's departure speech).
Among them, the six areas of artificial intelligence in the future of digital products and digital services may have an important impact. The authors list these six directions, explain their importance, the current application scenarios, and list the companies and research institutions that are in use.
Intensive learning
Reinforcement learning is a way of learning through experimentation and error, inspired by the process of learning new skills. In a typical reinforcement learning case, the agent takes action by observing the current state and maximizes the outcome of the long-term reward. Each time an action is performed, the agent receives feedback from the environment, so it can determine whether the effect of the action is positive or negative. In this process, the agent needs to balance the experience to find the best strategy and explore new strategies both in order to achieve the ultimate goal.
Google's DeepMind team uses reinforcement-learning techniques in Atari games and Go games. In the real world, reinforcement learning has been used to improve the energy efficiency of Google's data centers. Reinforced learning technology for the cooling system saves about 40% of energy consumption. Reinforcement learning has a very important advantage, its agents can generate a large number of training data at low cost. Compared to the supervised depth learning task, this advantage is very obvious, saving a large amount of manual labeling data costs.
Applications: including automatic driving of urban roads; navigation of three-dimensional environment; multiple agents interact and learn in the same environment, etc.
Generate the model
Unlike the discriminant model used to perform the classification and regression tasks, the generation model learns a probability distribution from the training samples. By sampling from the high-dimensional distribution, the model is generated to output new samples that are similar to the training samples. This also means that if the training data of the generated model is the set of images of the face, the model obtained after the training can also output a composite image similar to the face. Details can be found in the article by Ian Goodfellow. The structure of GAN is very hot in academe because it provides a new idea for unsupervised learning. GAN structure uses two neural networks: one is the generator, which is responsible for random input noise data into new content (such as synthetic images), the other is the discriminator, responsible for learning the real picture and determine the generator-generated Whether the content is real oneself. Confrontation training can be considered a type of game, the generator must be repeated learning random noise data with meaningful content, until the judge can not distinguish the authenticity of the synthetic content. The framework is being extended to many data patterns and tasks.
Application: simulation of time series features (eg, planning tasks in reinforcement learning); super-resolution images; restoration of three-dimensional structures from two-dimensional images; generalization of small-scale annotation datasets; prediction of the next frame of video; generation of natural language Of the dialogue content; art style migration; voice and music synthesis
Memory network
In order for artificial intelligence systems to be able to adapt to a wide variety of environments like humans, they must continually acquire new skills and remember how to apply them in future scenarios. Traditional neural network is difficult to master a series of learning tasks. This shortcoming is what scientists call catastrophic forgetting. The difficulty is that when a neural network training for A task, if re-training it to solve the B task, then the weight value of the network model is no longer applicable to the task A.
At present, there are some network structure can make the model have different levels of memory capacity. Including a short-term memory network (a recurrent neural network) that can process and predict time series; the DeepMind team of micro-neural computers, which combines neural networks and memory systems to facilitate learning from complex data structures; progressive neural networks , Which learns the lateral associations between individual models and extracts useful features from these existing network models to accomplish new tasks.
Applications: training agents able to adapt to new environments; robotic arm control tasks; autopilot vehicles; time series prediction (eg financial markets, video prediction); understanding of natural language and forecasting below.
Microdata learning micro model
All along, the depth learning model needs to accumulate a lot of training data in order to achieve the best results. For example, a team that only participated in the ImageNet Challenge used 1.2 million manually tagged image training models distributed across 1,000 categories. Leaving large-scale training data, depth learning model will not converge to the optimal value, can not be in voice recognition, machine translation and other complex tasks to achieve good results. The growth of data demand often occurs when a single neural network model is used to process the end-to-end case, such as inputting the original speech segment, and outputting the converted text content. This process works with multiple networks to handle one intermediate result (for example, raw voice input → phoneme → word → text output). If we want to use the artificial intelligence system to solve the task of training data scarcity, I hope the training model used in the sample as little as possible. When the training data set is small, over-fitting, outliers interference, training set and test set inconsistencies and other issues will follow. Another approach is to migrate models that have been trained on other tasks into new tasks. This approach is called migration learning.
A related problem is the use of fewer model parameters to build a smaller deep learning architecture, while the model's effect is to maintain the best. The advantage of this technology is more efficient distributed training process, because the training process needs to reduce the transmission parameters, and can easily be deployed in the model memory size limited embedded hardware.
Applications: Training shallow models to simulate deep network models trained on large-scale labeled training datasets; building model structures that have comparable but lesser parameters (eg, SqueezeNet); machine translation
Learning / reasoning hardware
One of the catalysts for the development of artificial intelligence is the evolution of the graphics processor (GPU), which, unlike the sequential execution mode of the CPU, supports large-scale parallel architectures that can handle multiple tasks simultaneously. Since the neural network must be trained with large-scale (and high-dimensional) datasets, the GPU is much more efficient than the CPU. This is why since the first GPU training in 2012, the neural network model - after the release of AlexNet, GPU has become a veritable gold-plated spade. NVIDIA continues to lead the industry in 2017, ahead of Intel, Qualcomm, AMD and the rising star of Google.
However, the GPU is not designed for model training or prediction, it was originally used for video game image rendering. GPU with high-precision computing capabilities, but suffered memory bandwidth and data throughput problems. This opens up new areas for large companies such as Google and many small start-ups that design and manufacture processing chips for high-dimensional machine learning tasks. The improvement of chip design includes more memory bandwidth. Graph calculation replaces vector computation (GPU) and vector computation (CPU), higher computation density, lower energy consumption. These improvements are exciting because they ultimately feed back to the user: faster and more efficient model training → a better user experience → more users to use the product → gather larger data sets → improve through the optimization model Product performance. As a result, systems that train and deploy models faster take a significant advantage.
Applications: Rapid training of models; Low-power predictive computing; Persistent monitoring of Internet of Things devices; Cloud service architecture; Autopilot vehicles; Robots
Simulation environment
As mentioned earlier, it is challenging to prepare the training data for the artificial intelligence system. Moreover, to apply an artificial intelligence system to real life, it must have applicability. Therefore, developing a digital environment to simulate the real physical world and behavior will provide us with the opportunity to test the adaptability of artificial intelligence systems. These environments present the original pixel to the artificial intelligence system, and then take some action based on the set goals. Training in these simulated environments can help us understand the learning principles of artificial intelligence systems, how to improve the system, but also provide us with the model can be applied to the real environment.
Applications: Simulation Driving; Industrial Design; Game Development; Smart City