Prospects of artificial intelligence The six areas need to pay close attention to

Source：Updated：2017-02-09 08:18:36

Nearly period of time, there are a lot of debate about artificial intelligence generally accepted definition. Some people think that artificial intelligence is the "cognitive computing" or "machine intelligence", while others get it confused with the concept of "machine learning". Especially some technology, artificial intelligence, however, is not it is actually a composed of multi-discipline broad areas, including robotics, machine learning, etc. The ultimate goal of artificial intelligence is to allow machines to replace human cognitive abilities need to complete tasks. To achieve this goal, the machine must be automatic learning ability, and not just follow the orders of the programmers. 
 
Prospects of artificial intelligence The six areas need to pay close attention to 
 
Artificial intelligence (ai) in the past ten years has made amazing progress, such as automatic driving, speech recognition and speech synthesis. In this context, artificial intelligence, this topic increasingly appear between colleagues and family gossip, artificial intelligence technology has penetrated into the nook and cranny of their lives. At the same time, the popular media also reported in almost every day and artificial intelligence technology giants, introduce their long-term strategy in the field of artificial intelligence. Some investors and entrepreneurs eager to learn how to value from this new field of mining, most people still racking their brains thinking what artificial intelligence will change. In addition, governments are struggling to cope with the effects of the automation brings to the society (e.g., President Obama's resignation speech). 
 
Among them, the six major areas of artificial intelligence in the future may be important influence on digital products and services. The author lists the six directions, explains the importance of their current application scenario, citing are using companies and research institutions. 
 
Reinforcement learning 
 
Reinforcement learning is a kind of learning method through experiment and error, it inspired by the human the process of learning new skills. In the case of the typical reinforcement learning, agents by observing the current state of the place, and then take action to make to maximize the results of long-term rewards. Every execution actions, agents will receive feedback from the environment, so it can determine the action is positive or negative effects. In the process, agents need to balance based on experience to find the best strategy and explore new strategy from two aspects, in order to achieve the ultimate goal. 
 
 
 
Google DeepMind team in Atari game and go against all use reinforcement learning techniques. In a real scenario, the reinforcement learning has been used to increase the energy utilization in Google data center. Reinforcement learning technology for the cooling system saves about 40% of the energy consumption. Reinforcement learning has a very important advantage, its agents at low cost simulation can generate a large number of training data. Compared with the depth of the supervised learning tasks, this advantage is very obvious, save the cost of a large amount of manual annotation data. 
 
Application: includes the autopilot of city road; The three dimensional environment navigation; Multiple agents in the same environment interaction and learning 
 
Leading researchers: Pieter abbeel (OpenAI), David Silver, cholesterol-conscious patrons DE Freitas, Raia Hadsell (Google DeepMind), Carl Rasmussen (Cambridge), Rich Sutton (Alberta), John Shawe - Taylor (UCL), etc 
 
Technology on behalf of the company: Google DeepMind, Prowler. IO, Osaro, MicroPSI, Maluuba/Microsoft, NVIDIA, Mobileye, etc 
 
Generate models 
 
Different from used to complete the task classification and regression discriminant model of generation model learned from the training sample a probability distribution. By sampling from the distribution of high dimensional and generate the model output is similar to the training sample of the new samples. This also means that if the generation model of training data is facial image set, then the model can also output resulting from the training similar to the synthesis of face images. Details you can refer to Ian Goodfellow. He proposed to generate counter model (GAN) structure present in academia is very hot, because it gives unsupervised learning provides a new way of thinking. GAN structure with the help of two neural networks: one is the generator, which synthesize the noise of the random input data into new content (such as synthetic image), the other is a discriminant, responsible for learning the real pictures and judge whether the content of the generator to generate real ones. Training can be considered to be a kind of game, the generator must be repeated learning with random noise data synthesis of meaningful content, until the discriminant can't distinguish synthetic content authenticity. This framework is being extended to many data mode and tasks. 
 
Application: the characteristics of the simulation time series (for example, in reinforcement learning planning tasks); Super-resolution image; From the two-dimensional image restoration three-dimensional structure; Generalization of small-scale annotation data set; Forecasting the next frame of video; Generate natural language dialogue content; Art style migration; Voice and music synthesis 
 
Technology on behalf of the company: Twitter Cortex, Adobe, Apple, Prisma, Jukedeck, Creative. Ai, Gluru, Mapillary, Unbabel 
 
Leading researchers: Ian Goodfellow (OpenAI), Yann LeCun and Soumith Chintala (Facebook artificial intelligence research institute), Shakir, Mohamed and A? RON van den Oord (Google DeepMind) and so on 
 
Memory network 
 
To make artificial intelligence systems like humans can adapt to various environment, they must continuously learn new skills, and remember how to apply these skills in the scene in the future. The traditional neural network is very difficult to master a series of learning tasks. Scientists call this drawback is disastrous. The difficulty is that when A neural network for A task to complete after the training, training again if it B tasks, the network model of weight value is no longer suitable for task A. 
 
At present, there are some network structure could make models have different levels of memory capacity. Including both short-term and long-term memory network () a recursive neural network can handle and forecasting of time series; DeepMind team micro computer nerve, it is a combination of neural network and memory system, so as to learn from complex data structures; Progressive neural network, it study of lateral connection between individual model, from the existing network model to extract useful features, used to complete the new task. 
 
Application: training to adapt to the new environment of agents; Robot arm control task; Automatic vehicle; Time series forecasting (such as financial markets, video prediction); To understand natural language and forecast below. 
 
Technology on behalf of the company: Google DeepMind NNaisense, SwiftKey/Microsoft Research. 
 
Leading researchers: Alex Graves, Raia Hadsell, Koray Kavukcuoglu (Google DeepMind), Jurgen Schmidhuber (IDSAI), Geoffrey Hinton (Google Brain/Toronto) 
 
Micro data to study micro model 
 
Has long been a deep learning model is the need to accumulate a large number of training data to achieve the best effect. ImageNet challenge, for example, a cat show in team with 1.2 million copies in 1000 categories of manual annotation image training model. Leave the large-scale training data, deep learning model will not converge to the optimal value, nor in complex tasks such as speech recognition, machine translation to get good effect. Data volume demand growth often occur with a single neural network model processing end-to-end cases, such as input of the original speech segment, required output text after conversion. This process with multiple network work together each processing step intermediate results different (for example, the original voice input to phonemes, words and text output). If we want to use artificial intelligence system to solve the task of training data scarce, hope model training samples used as little as possible. When the training data set is small, through the interference fitting, abnormal values, inconsistent training set and testing set distribution problems will follow. Another way is to the trained model on other task migration to the new task, this method is called migration. 
 
A related problem is to use less of the model parameters less deep study of architecture, and the effect of the model is best. This technique has the advantage of more efficient distributed training process, because in the process of training need to reduce the transmission of parameters, and can be easily will be deployed in the memory size restricted model embedded hardware. 
 
Application: a training model to simulate on large-scale labeled training data set obtained the depth of the network training model; Build effect quite but less model structure parameters (such as SqueezeNet); Machine translation 
 
Technology on behalf of the Company: Geometric Intelligence/Uber, DeepScale. Ai, Microsoft Research, Curious ai Company, Google, Bloomsbury ai 
 
Leading researchers: Zoubin Ghahramani (Cambridge), Yoshua Bengio (Montreal), Josh Tenenbaum (MIT), Brendan Lake (NYU), Oriol Vinyals (Google DeepMind), Sebastian Riedel (UCL) 
 
Studying/reasoning hardware 
 
One of the catalyst to promote the development of artificial intelligence is updating the graphics processor (GPU), different from the sequential pattern of the CPU, GPU support massively parallel architecture, can handle multiple tasks at the same time. In view of the neural network must use mass (and high dimension) training data set, the efficiency of the GPU is far higher than the CPU. This is why since 2012, when the first GPU training neural network model - AlexNet released after the GPU has become a veritable gold mining shovels. NVIDIA to continue leading industry in 2017, ahead of Intel, Qualcomm, AMD and upstart Google. 
 
However, the GPU is not specially designed for training or prediction, it was originally used for image rendering video games. GPU has the ability of high precision calculation, but a memory bandwidth and data throughput. For the large company and many small entrepreneurial firms such as Google has opened up new areas, they are for high dimensional design and manufacture of processing chip machine learning task. The improvement of chip design points include larger memory bandwidth, figure calculation instead of the vector (GPU) and vector calculation (CPU), the calculation of higher density, lower energy consumption. These improvements exciting because eventually and feedback to the user's body: faster and more effective model of training to better user experience, the use of user more products to collect more data sets, through optimization model to improve the performance of the product. Therefore, the training and deployment model faster system occupy significant advantages. 
 
Application: the training of the model; Low energy consumption prediction calculations; Continuous monitoring iot device; Cloud service architecture; Automatic vehicle; The robot 
 
Technology on behalf of the company: Graphcore Cerebras, Isocline Engineering, Google (TPU), NVIDIA (DGX - 1), Nervana Systems (Intel), Movidius (Intel), Scortex 
 
The simulation environment 
 
As mentioned earlier, training data for artificial intelligence system is very challenging. Also, if you want to apply artificial intelligence systems to real life, it must have applicability. Therefore, development of digital environment to simulate the real physical world and the behavior will provide us with system suitability test of artificial intelligence. The environment to the artificial intelligence system to present the original pixels, then according to the target and take some action. Training in the simulation environment can help us to understand the learning principle of artificial intelligence system, how to improve the system, also provide us with the model can be applied to the real environment. 
 
Application: driving simulation; Industrial design; The game development. Wisdom city 
 
Technology company representative: Improbable, Unity 3 d, Microsoft (Minecraft), Google DeepMind/Blizzard, OpenAI, Comma. Ai, Unreal Engine, Amazon Lumberyard 
 
Leading researchers: Andrea Vedaldi (Oxford)