Google's AlphaGo beat top professional go players, the symbol of the machine finally conquered the human, but the evolution of deep learning (Deep Learning) technology does not end with this game, such as Baidu's voice recognition rate has been raised from 89% to 99 In addition, the number of jobs related to deep learning technology was almost zero in 2014 and has now risen to around 41,000.
Deep Learning is currently in Peak of Inflated Expects (Figure 1) in Gartner's Hype Cycle (Figure 1). One of Gartner's five major predictions for 2017 is that Deep Learning will move from the edge Technology evolves into one of the key elements of analytic technology.
Figure 1 Gartner emerging technology development cycle, the depth of learning is currently in the expected expansion period.
Deep learning is a change in machine learning, that is, solving business problems through data retrieval. Deep Learning searches for Intermediate Representation, which enables standard machine learning techniques to be further exploited. More complex issues may also be used to solve other problems, not only with higher accuracy, fewer observations, fewer cumbersome manual fine-tuning tasks.
With deep learning, computer models can feed in a wealth of complex data such as images, speech and text, etc. For example, deep learning algorithms can analyze retinal scan results and "understand" what patterns represent healthy or diseased retinas And shows the name of a particular disease.) This "understanding" of the process must rely on brute-force high-performance computing, and to some extent, it can eliminate outdated manual preparation of functional data such monotonous jobs.
Figure 2 illustrates how to improve the visual abstraction of automatic search when face recognition is based on deep learning techniques.
The most common category of deep learning is the Feedforward Deep Neural Network, which uses the innumerable layers of interconnected processing units to "find" the proper mediation code from raw input data. Deep Neural Networks provide a Powerful architecture for a wide range of business issues.
To train a deep neural network with thousands or even millions of parameters, there must be a highly iterative computationally intensive program that takes advantage of the "Heuristic" concepts of "Gradient Descent" and "Backpropagation" Numerical Optimization Tips These optimization tips are now only available on such a large scale with breakthroughs in high-performance graphics processor (GPU) architectures in recent years.
Recently, deep learning has led to a number of major scientific breakthroughs and is considered one of the key players in artificial intelligence (AI) and will impact most industries in the next three to five years. Therefore, data and analytics executives should now take action to understand relevant Technical advantages and challenges.
Deep learning inherits all the advantages of machine learning and its greatest potential lies in the ability to learn the mediation codes in a specific area to enhance the effectiveness of the resulting solution.However, the application of deep learning techniques is a real risk, and will Increased risk due to inappropriate data, lack of transparency in the model, lack of relevant data science and programming skills, the need for high-performance computing infrastructures, and uncertain or inconsistent administrative support.
Fortunately, most of the deep learning capabilities will be introduced into the enterprise organization in a user-friendly form through Application Programming Interfaces (APIs) or Software as a Service (SaaS), as well as packaged applications.
Deep Learning will continue to evolve and is currently the most preferred technology in the Predictive Analytics area to handle the types of data previously considered un-traceable or machine-learnable such as video, voice and video, Data fusion, it is also more accurate than other techniques, and because of its data fusion capabilities over other machine learning methods, Gartner predicts that in 2022, deep learning will dominate big data analytics.
The world of data analytics appears to have immense potential, but data integration remains a challenge. Deep learning, automated tools such as Natural-Language Generation (NLG) work together with the right data to work seamlessly, but if Data is not easy to integrate, you need professional data integration and scientists, in order to effectively use these new tools.
Peter Krensky, a senior research analyst at Gartner, said: "In the future, there will be a combination of professional data integrators and data scientists who use relevant technologies to improve efficiency while also having a small group of citizen data scientists and citizen data integrators serving as either formal or semi-official Formal roles Data and analytics executives should enable multi-purpose teams and use sandboxes to reduce the risk for less skilled employees.
Learning Deep Prediction Analysis Technology Learning
By 2018, 80% of data scientists will include deep learning (deep neural networks) as one of the standard toolkits, and deeper learning is gaining in popularity both at the project and hiring levels. The rapid evolution is due in part to investment-related research in large research labs such as Facebook and IBM, where about 30% of data science platform vendors have introduced products using the first version of deep learning technology.
However, deep learning is not an independent technology. In fact, it is one of the third wave analysis techniques developed by machine learning. Enterprises should consider machine learning as a potential service project, and its possible applications Including anomaly detection, voice control and query, emotion analysis and face recognition.
In the coming years, deep learning will drastically impact analytics and operational practices, and users of smartphones are already seeing significant changes at the moment, such as the voice recognition capabilities of Apple's Siri and Google Voice, deep learning of new technologies It has become an actual function close to daily life. In various fields such as image recognition, phonetic transcription and machine translation, new breakthroughs have been made almost every week.
Gartner estimates that there are currently more than 2,000 vendors that begin offering deep learning tools, cloud services, application interfaces, packaged applications and consulting services, all from start-ups to technology giants, so virtually any organization in the industry has There are hundreds of deep learning products to choose from, while data and analytics executives also gain a unique opportunity to evaluate how deep learning can help digital businesses from a practical perspective.Gartner predicts that by 2019, deep learning will Become one of the top handlers for Best-in-class failure prediction, demand forecasting and fraud prediction.
In-depth learning in recent years, several breakthroughs in the field of cognition are:
Image recognition
At the end of 2015 and early 2016, Microsoft's ResNet and Google's GoogLeNet (V4) demonstrated a stunning image recognition system that outperformed humanity in the work of ImageNet Image Classification.
· machine translation
Google has introduced Google Neural Machine Translation (GNMT), claiming to have significantly improved the most advanced in-machine translation in the past.
· Speech Recognition
For similar work, the performance of Baidu's voice-to-text service has surpassed that of mankind, and the further development that is expected to take place may lie in the non-cognitive field where deep learning is gradually marching:
Fraud detection
PayPal has begun using deep learning as the best-of-breed best way to intercept fraud payments.
· Recommended system
Amazon has recommended deep learning to best-of-breed products of its kind.
From now until 2019, companies will apply deep learning technologies to the consumer world, primarily through integrated suite of business applications, appliances or application programming interfaces Data and analytics executives should start looking for opportunities to bring deep learning to the organization, especially Important business issues with a clear "perceived component" should involve academics, research labs or consultants, learn more about in-depth learning and consider potential new start-ups to acquire relevant talent and skills.
Business Intelligence and Analytics will be Included in NLG
By 2019, 90% of modern business intelligence and analytics platforms will include natural language generation as one of the standard features
Data visualization has been one of the key drivers of modern business intelligence (BI), but sometimes it is difficult to fully interpret this form of data.Natural language generation can create content-based text or voice narratives for data research results, Produce complete stories for key projects through visualization. Today's BI teams integrate a standalone natural language generation engine, but this will change as technology evolves; natural language generation will enable next-generation business intelligence and analytics The platform automatically finds, visualizes and provides the key findings that will broaden the audience for analytics while reducing the time and cost associated with regular volume reporting.
Data and analytics executives should begin integrating natural language generation techniques with established business intelligence / data searches or other tools while monitoring their potential and growth for business intelligence, data exploration and data science, and the newly emerging start-ups path.
Finally, Gartner suggests the following for data and analytics executives in charge of analytics technology and business intelligence modernization: First, revisit the inherent problems of lacking solutions but still having useful data; in some of the more familiar algorithms, deep learning related Pilot programs are beginning to emerge and surpass other technologies, and it is advisable to start with those issues that have already reached consensus on the expected results.
Second, the easiest solution to adopt; for many business problems, "Shallow" machine learning will still be the best way to avoid the inherent complexities of deep learning technology.
In addition, a major pool of talent must be established before beginning a formal, deep learning experiment, and if there is a shortage of talent, consider outsourcing through service providers or academia.
Last but not least, validate the results of deep learning immediately, closely monitor fairness, compliance with laws and regulations and ethics, but do not abandon experimenting plans for deep learning too soon.