Depth | Achieving These Values â€‹â€‹$400 Million: See How Nervana Used Deep Learning to Achieve Data Revolution

In August 2016, Intel acquired Nervana, a machine learning startup, for $400 million in order to strengthen its capabilities in the field of artificial intelligence. Only two years after the start-up company was established, it was recognized as the leader in machine learning technology development. Dr. Naveen Rao, co-founder and chief executive officer of Nervana, recently discussed the topic of deep learning with everyone at StrataHadoop. Through its PPT, we can also understand the value of Nervana.

Naveen explores many aspects, including the benefits of deep learning relative to other machine learning technologies, the latest advances in the field, in-depth learning workflows, the challenges of developing and deploying deep learning solutions, and the standardization of building and extending deep learning solutions Tools and more.

Different from everyone's past understanding, what exactly is deep learning? This is a machine learning method that mimics the human brain's mechanism for analyzing data, grabbing features from multiple levels of abstraction. What we think is really important is to extract useful information from the data and make judgments based on statistical information. We have authoritative experts in the field of data science, which enhances learning performance through a large amount of data and ultimately achieves a high degree of visualization.

This ability is very powerful, similar to the significance of computers for humans in the past twenty to twenty-five years. In just two years, this new method of deep learning has been widely used in industry, including communication, voice processing, image recognition, video processing, etc. A large amount of data can be used by people, so I think The future prospects are very bright.

There are many models for deep learning, but the basic structure is different and may lead to different characteristics of these models. Here I will quickly review these common depth models with you.

The model in the upper left corner is the most commonly used convolutional neural network (CNN) model and is well-suited for visual systems and image analysis. The model in the upper right corner is the Recurrent Neural Network (RNN) which is suitable for various time or sequence based modeling. Financial systems and language models often use RNNs. There are many other non-mainstream methods, such as the following three: stacked automatic encoders, multilayer perceptrons (MLPs), and deep trust networks (DBNs). Many innovations in the coming five years will come from the field of stacked automatic encoders, but at present we do not know how to start selecting targets.

Here we can see some real experimental results. This system abstracts the entities and extracts rich representative features. This is our true source of strength, and its anti-interference ability is very good and cannot be easily changed.

The significance of this is that you can automatically handle tasks that previously required manual completion. We also did some tests so that everyone can see their performance. At present, trained humans usually have a 5% error rate, and deep learning has only 3% error rate in image and speech tasks. Therefore, we believe that in a few years time, the ability of deep learning in these two areas will even go far beyond humanity.

This is a deep credit network, and we see how it can be used to solve data problems. There are two ways to use this system: The first is that we can enter handwritten patterns corresponding to some numbers, the system can use these data samples to determine which corresponding number; the second is that we give a certain number, the system can simulate It corresponds to a variety of possible handwriting styles.

Here, for example, the system judges that this is the number "5" according to a series of "5" handwritings.

For another example here, when we enter the number â€œ0â€, the system goes through a series of handwriting style processing, and it will display various â€œ0â€ patterns in real time. However, no doubt, we can still visually recognize that this is a number. "0".

This is a very interesting thing, which means that we can input a series of examples from which to abstract out specific common features.

This deep learning platform implements a 3D convolutional model for behavior detection. Based on a common data set of 100 categories and more than 13,000 videos, the training speed is approximately 3 times faster than the competition's framework. At the same time, the platform can also be extended to other scenes, object recognition, parallel comparison of behavioral similarity, video retrieval and anomaly detection.

Potential applications include: implementation of safety monitoring, traffic control and vehicle management, air traffic control detection, security system retrieval based on facial recognition and image processing, awareness and collision avoidance system for automatic driving in passenger-intensive places such as airports or subway stations. , baggage testing in public places, etc.

Voice can be seen as a random combination of words, so it is difficult to convert speech into text. However, after a lot of training, the system can recognize most of the words. The performance of deep learning in speech-to-text conversion is also impressive. Companies such as Baidu and others have very mature natural language processing technologies that can be converted into any language as needed. This is magical as magic.

As an inevitable law of historical development, when a certain inflection point is reached, it will suddenly erupt, and now it is at the inflection point of data science, using deep learning to maximize the use of data.

The CPU's training time is more than thirty times that of a single GPU.

Data parallelism is one of the most commonly used methods. There is a full-depth network in each processor, and the parameters in each data container are uniformly coordinated to the parameter server. But this is not the best way.

A better way is to model parallel computing as shown.

Another advantage of Nervana is the I/O range. The more processors in general, the faster the deep learning. However, with the increase in the number of processors, common industrial systems will reach a certain limit without increasing their learning speed. The Nervana platform can not only improve the learning speed of a single processor, but also does not have the upper limit of learning speed, and can increase the number of processors as needed.

We are still working hard to develop new technologies and strive to increase our current speed by more than ten times next year.

The Nervana platform is a full-stack solution based on the Nervana Deep Learning Framework platform and the Nervana Cloud for input, construction, training and deployment.

As the core competitiveness of Nervana, Deep Learning has built features such as image classification, target positioning, video retrieval, text analysis, and machine translation.

Nervana has the fastest deep learning library.

Nervana's Python Deep Learning Library is user-friendly, extensible, supports multiple deep learning models, and provides interfaces to Nervana Cloud. It also supports multiple backgrounds (including Nervana Engine, GPU, and CPU).

This is the system's web interface. Nervana provides users with a large number of APIs that can be called directly.

The role of deep learning is to create a framework for finding useful information in the data, but there are still many difficulties in making this framework faster, processing larger, and covering a wider range of solutions.

Nervana currently has the most advanced deep learning platform, making it easy to use the relevant tools developed to abstract representative target features from complex relationships. In addition to the various applications mentioned earlier, it can also be used to quickly locate oil wells, natural gas fields, and agricultural refinement operations.

Via NextBigFuture