Four deep beasts under the deep study seat
Nowadays, more and more people are willing to talk with their virtual personal assistants. Just move their mouths and let Siri/Alexa/Rokid help you to do things like WeChat, booking tickets, setting alarm clocks, and remind you to take medicines. Meeting, how can such an intimate baby jacket that does not need to pay salary be loved? Virtual assistants are step by step closer to real-life personal assistants, and it is the deep learning technology behind them. In addition to virtual assistants, deep learning technology will also be the core technology in many fields such as computer vision, autopilot, and voice recognition. The four key elements depth study and practice of: computing power, algorithms, data and application scenarios, like the four law enforcement animal as to ensure the practical application and depth of learning, indispensable.
Deep learning is a technique in which the neural network of not less than two hidden layers performs nonlinear transformation on the input or represents learning, and constructs a deep neural network to perform various analysis activities. The deep neural network consists of one input layer, several hidden layers, and one output layer. There are several neurons in each layer, and there are connection weights among the neurons. Each neuron mimics the nerve cells of the organism, and the junction between the nodes simulates the connection between nerve cells. To sum up is this:
A special property of this kind of flow graph is depth: the length of the longest path from one input to one output. Deep learning is not a new concept, but it was led by Hinton et al. in 2006. However, although many people are talking about deep learning in recent years, what are the pitfalls of this technology in practical use? What are the elements needed to start a mature company that relies on deep learning? The following is our view.
| Computing power
First of all, deep neural networks are complex, and there are many training data and large amounts of computation. Deep neural networks have many neurons, and the number of connections between neurons is quite alarming. From a mathematical point of view, each neuron must contain mathematical calculations (such as Sigmoid, ReLU, or Softmax functions), and the amount of parameters that need to be estimated is also enormous. In speech recognition and image recognition applications, there are tens of thousands of neurons, tens of millions of parameters, and the complexity of the model leads to a large amount of computation. So computing power is the foundation of deep learning applications .
Not only that, but also computing power is a powerful tool to promote deep learning. The more computing power, the more experience accumulated and the faster the iteration speed . Baidu's chief scientist Dr. Wu Enda believes that the frontier of deep learning is shifting to high-performance computing ( HPC), which is also one of his current focuses in Baidu, Dr. Wu believes that a lot of success in deep learning is due to the active pursuit of available computing power. In 2011 Jeff Dean (Google's second-generation artificial intelligence learning system Tensorflow's One of the designers) founded and led the Google Deep Learning Group, using Google Cloud to expand deep learning; this has enabled deep learning to be pushed into the industry. In 2013, Dr. Coates and others established the first HPC-style deep learning system, which has increased scalability by 1-2 orders of magnitude, and has made a profound progress in deep learning— computational ability for deep learning. The support and promotion are irreplaceable .
At present, the leading technologies in this area are still large Internet companies like Baidu and Google. Of course, there are some startup companies like Horizon Robotics that have achieved success in this area. The horizon was founded by Dr. Yu Kai, the head of Baidu Deep Learning Institute. The deep neural network chip designed by the robot company can support the tasks of image, voice, text, control, etc. in the deep neural network rather than the traditional CPU chip compared to the traditional CPU chip. This is more efficient than using software on the CPU, and it will improve 2-3 orders of magnitude .
| Algorithm
At a time when computing power has become increasingly cheap, deep learning attempts to create much larger and more complex neural networks. We can think of algorithms as deep learning neural networks or computational thinking. The more complex this neural network is, The more accurate the signal is, the more common algorithms currently include Deep Belief Networks, Convolutional Neural Networks, Restricted Boltzmann Machines, and Stacked Auto. -encoders) The supervised learning methods represented by deep convolutional neural networks are currently the most effective and most commonly used.
However, the current problem is that everyone puts the degree of focus on data and operations, because the differences in the neural network itself will not be great, and the core algorithm of the neural network is too difficult to raise , and still faces problems like local optimality, cost function and the entire nerve. Network system design and other issues, but this also gives many startup companies a new idea, why not go against it and avoid the “single wood bridge†that carries thousands of troops. If we can optimize the algorithm, the future is not limited.
| Data
Now that deep learning is rapidly becoming a hot topic in the field of advanced data analysis, the absolute amount of data is a key factor in promoting the development of deep learning tools and technologies . According to Daniel McDuff, chief scientist and research director at Affectiva , technology can only work better when new companies have accumulated enough data . For those deep-learning applications, not only does it require a lot of data training and improvement during development, it also needs real-time user data after product promotion, and iteratively updates.
China still has a great opportunity for deep learning competition. Available Internet data and low-cost crowdsourcing will bring huge amounts of data and very low data labeling costs to Chinese companies. However, the problem faced by the domestic market is that a large amount of data is controlled by Internet giants such as BAT. It is very difficult for start-up companies to obtain data to improve and update the neural network for deep learning. Especially after the product is launched, it may also face the malicious exclusion of large companies. It's even harder to get data, and it's almost impossible to survive in the cracks.
| Application scenarios
There are not many scenarios currently used for deep learning technology. The most common and successful areas are the two application scenarios of speech recognition and image processing . The three major beasts mentioned above—computing power, algorithms, and data belong to the development side. The scene is at the consumer end level. With the continuous development of deep learning technologies and the increasing demand of users, more and more application scenarios for deep learning will occur . For example, many smart phones have built-in face recognition to classify photos. Can achieve quite accurate rates; Alipay and other financial tools are also likely to use face recognition to improve security... The future of deep learning is not limited to only two areas of speech recognition and image recognition, there are more possibilities . For those start-ups, it is better to compete with a large company that has more than a decade of data precipitation, such as Google, Facebook, Amazon, and BAT, in order to develop its own small world.
Nowadays, the intensity of deep learning is not weaker than any other field. Internet giants are trying to split this piece of cake. In fact, they want to do deep learning, computing ability, algorithms, data, and application scenarios. , and BAT and other giants in these areas all take advantage of resources, it is difficult for start-up companies to take into account four points , especially in terms of data, so use their own comparative advantage to seize one point to innovate, regardless of computing power, The algorithm is still an application scenario. As long as there are innovations, it will help you to take the initiative in the market.
Lei Feng Net Note: This article was originally published in WeChat public number linear capital (public number: LinearVenture), authorized Lei Feng network (search "Lei Feng network" public number concerned) release. Reprint please contact us for authorization, and keep the source and the author, not to delete the content. The official public platform for linear capital focuses on early investments in Pan-Smart, Fintech, and VR/AR.
These products are used in intelligent high-frequency switching power supply, electric power supply, railway power supply, large LED power supply, and uninterruptible power supply(UPS)
through-type power connector,durable,reliable,power transfer,efficient
Huizhou Fibercan Industrial Co.Ltd , https://www.fibercaniot.com