How many buses can I take to Guilin southwest central school

1.

301 road → 99 road 24.6km

Vocational Ecation Center

walk about 790m to Lingchuan No.2 Middle School Station

< P >

take No.301 Road, pass 23 stops, and arrive at Xinjian intersection station

walk about 30m, Arrive at Xinjian intersection station

take No.99 bus, pass 17 stops, arrive at Beidou business district station

walk about 410 meters to Guilin No.8 Middle School

2. T is ton, 1 ton = 2000 Jin = 1000 kg

3. It's 0.14ct, right? It means the weight of the diamond is 0.14 carat.

4.

NVIDIA GPU, amd GPU or Intel Xeon Phi

1. it's easy to build CUDA's deep learning library with NVIDIA's standard library, but AMD's OpenCL's standard library is not so powerful. Moreover, CUDA's GPU computing or general GPU community is large, while OpenCL's is small. It is more convenient to find good open source methods and reliable programming suggestions from CUDA community. NVIDIA began to invest in deep learning from the beginning, and the return is quite good. Although other companies also invest in deep learning, they start late and lag behind a lot. If we use other software and hardware besides nvidia-cuda in deep learning, we will take a detour

2. Intel's Xeon Phi supports standard C code, and it's easy to modify these codes to accelerate on Xeon phi. This function sounds interesting. But in fact, it only supports a small part of C code, which is not practical. Even if it is supported, the implementation is slow. Tim has used 500 Xeon Phi clusters and encountered one after another pitfalls. For example, Xeon Phi MKL and python numpy are incompatible, so he can't do unit testing. Because Intel Xeon Phi compiler can't rece the code of template correctly, such as switch statement, a large part of the code needs to be refactored. Because the Xeon Phi compiler does not support some C + + 11 functions, the C interface of the program should be modified. It's troublesome, it takes time, it's maddening. Implementation is also slow. When the tenor size changes continuously, we don't know whether it is a bug or thread scheling that affects the performance. For example, if the size of FC or dropout is different, Xeon phi is slower than CPU

< / OL >

the fastest GPU within budget

what is the high speed of GPU for deep learning? CUDA core? Clock speed? Or ram size? These are not. The most important factor affecting the performance of deep learning is the bandwidth of video memory. GPU's video memory bandwidth is optimized at the expense of access time (delay). On the contrary, the CPU uses less memory and has faster calculation speed, such as the multiplication of several numbers (3 * 6 * 9); For example, matrix multiplication (a * b * c) is used. GPU with its video memory bandwidth is good at solving the problem of large memory

therefore, when buying a fast GPU, first look at the bandwidth

5. Mule intuitively affirmed that his performance this season is also higher than that of W card. In fact, it should be so.
it can be said that the 14T mule has reached a new high again. Of course, the price may also be in a very high state.

at present, the Hanfu has just come out at 8500w. I think it may be around 3 to 4000W

6. It's not easy for AWS to give up its own wheel and make it clear that it will support a system mainly developed by the open source community. For AWS, the most important concern is the user experience, and then it is to buy resources to make money. The safest thing here is to support all the popular DL frameworks. However, AWS can emphasize that it will invest heavily in mxnet in the future, mainly affirming the development prospect of the framework and the work of partners For example, when I learned that Amazon CTO blogged this morning, my first reaction was who is CTO? Why haven't I heard of it?)

mxnet was an interest project that several people started with pure enthusiasm for technology and development. They did not expect to graate or make money with it. To be able to expand step by step, I think the most important thing is the investment of each small partner in this matter and the mission of lowering the threshold of deep learning. It also allows us to only care about the "amount of data and computation", rather than how to use the hardware effectively; You just need to "write the mathematical formula well, and you don't want to know which layers you support and what they do", regardless of how to train the automatic derivation; Just "hand over the data to the cloud, and then spend money to rent computing power", rather than how to manage and optimize the cloud.

7. Although both inside and outside the instry can feel the hot wave of artificial intelligence in recent years, in fact, artificial intelligence technology does not appear in recent years. Since the 1950s and 1960s, artificial intelligence algorithms and technologies have been hot for a time. With the development of time, they continue to evolve and evolve, and have experienced a process from heat to decay
in recent years, artificial intelligence has made everyone feel its very hot and sustainable development. Therefore, we believe that this round of rapid development of artificial intelligence benefits from the rapid development of IT technology over the years, which brings computing power and computing distance to artificial intelligence, so as to provide support for artificial intelligence algorithms
the R & D of artificial intelligence technology and various artificial intelligence applications of enterprises are constantly landing, which directly promotes the rapid development of the overall artificial intelligence instry. The overall scale of the core instry of artificial intelligence is close to 100 billion yuan, which can be said to be one of the instries with huge scale. Moreover, from the perspective of future development trend, it is estimated that this year, the overall market scale will reach 160 billion yuan, so the growth rate is still very fast. Today, in some cases, machines trained through deep learning perform better than humans in image recognition, including looking for cats and identifying cancer signs in the blood. Google's alphago learned go and trained a lot for the game: constantly playing against itself
what is deep learning
deep learning is a technology to realize machine learning. Early machine learning researchers also developed an algorithm called artificial neural network, but it was unknown for decades after its invention. Neural networks are inspired by the human brain: the interconnection of neurons. However, the neurons in human brain can connect with any neurons in a specific range, and the data transmission in artificial neural network has to go through different layers and different directions.

8. Deep learning algorithm can be used to analyze and judge the characteristics of human expression and voice, and multi-modal analysis mode can be used.

9.

Deep learning is an algorithm revolution, which brings the rapid development and application of artificial intelligence, solves the processing of video, image, sound, language and text, and achieves the level of human recognition or cognition of objects to a certain extent

deep learning is only a sub field of machine learning, which is an algorithm of artificial neural network inspired by the structure and function of brain. Deep learning just needs a very large neural network to train more data, and needs more powerful computer and computing power

if we build larger neural networks (more hidden layers 10-100, or even more) and train to feed more and more data to the model, the performance of deep learning will continue to improve. This is different from other traditional machine learning algorithms. Deep learning technology will reach a new level in performance

hope AI can change the world again

10. Well, there is no accurate answer for the moment, because the main network is not online, and the mining parameters are only estimated by installing some known information. According to the existing information, it is estimated that 350000-500000 fils will be issued on the first day of the main network.