This article was awarded by the cooperative media mashdigi Reprinted with authorization.
Aiming at the artificial intelligence systems currently widely used in Google Search, Google translation, Google Photos, speech recognition and other services, Google not only explained that the TPU can reduce the learning and training time through the customized processor TPU, but also further explained the neural network composition learning behind it. The principle of model design makes artificial intelligence systems more “smart” by accumulating new learning experiences from the continuous learning process.
Quoc Le, a research scientist who belongs to the Google Brain team, explained that in order to simulate the thinking mode generated by the connection of human brain neurons, Google’s internal research team used a layered neural network to overlap 1010 Group (10 billion groups) network connection and combination, and through the evolutionary algorithm and reinforcement learning algorithm, the system can accumulate experience from the learning process, and then achieve the effect of learning from the learning experience, just as humans would learn from the rules of experience The situation of learning different ideas.
The current method adopted by Google is mainly to use the master neural network to derive the subset learning model architecture, and use it to perform specific mode learning and training, while evaluating the final training results, and finally pass the training results back to the master neural network. In the process of network, it will be judged whether to modify the data according to the learning situation.
The new learning model has been used to support
CIFAR-10 image recognition, and for natural language processing
Penn Treebank data set language model, and as the application basis for many Google services, such as a faster way to identify image content as “panda”, and real-time display of the next candidate vocabulary in the input method.
Artificial intelligence has been in the laboratory stage from a few years ago, and a large amount of it has entered ordinary life in a short time. Almost applications such as mobile phones and Internet services contain artificial intelligence technology components and bring higher computing efficiency than traditional ones. Therefore, It has become a new development model of computer computing. At the same time, because more and more data are generated every day, the traditional calculation model is no longer effective. Therefore, the analysis and calculation model of artificial intelligence will become the mainstream in the future. After Google made the TensorFlow learning model open source, the current learning model is also widely used in a short period of time. Almost most artificial intelligence systems use the TensorFlow framework for deep learning training.
However, although the new design allows the artificial intelligence system to learn in a more efficient manner, and even generates new learning experience rules on its own, it still cannot let the system itself understand the nature of the current processing content, that is, it cannot have emotions like humans. Therefore, it is impossible to judge whether the information is wrong in the learning process. It can only rely on the continuous learning process to find errors and re-correct them, or make adjustments through artificial methods.
In this way, by knowing how the system calculates, and then master how the data is used and calculated, avoiding the phenomenon of “overrun” in the system, allowing users to make good use of the artificial intelligence calculation model to bring a more convenient and efficient user experience At the same time, because the artificial intelligence system is still unable to judge the nature of the content on its own, even if the artificial intelligence technology develops at a fairly rapid angle, the “people” who judge and use as the final result still play an important role.
Using artificial intelligence to promote a new medical system
In addition to applying artificial intelligence systems to its services, Google also explained that more and more technological applications are closely related to artificial intelligence, such as the use of artificial intelligence to analyze and predict various diseases, including assisting clinicians in detecting breast cancer in the lymph nodes.
Metastasis, or screening and judging diabetic retinopathy, and at the same time, the artificial intelligence system is further combined with various medical auxiliary equipment to achieve the early prevention effect of various diseases, and even further assist in determining whether the patient needs hospital observation in the future, and Judging future illnesses that will affect your health, and at the same time, you can gain insight into the patient’s possible needs based on case data.
At present, Google has cooperated with medical researchers such as the University of California San Francisco School of Medicine, Stanford University School of Medicine, University of Chicago School of Medicine and other medical researchers to conceive how to combine machine learning technology and clinical diagnosis expertise to improve medical effectiveness and reduce unnecessary extras Medical expenses and medical negligence, while assisting clinicians to use more accurate analysis and judgment to make better diagnosis results. Even for remote areas or developing countries where medical human resources are relatively scarce, the combination of cloud analysis and computing resources can also help physicians make more accurate medical diagnoses.
However, before implementing this development, Google is also facing the difficulty of data access caused by the differences in data record formats adopted by different medical institutions. Therefore, it will now adopt the medical open data standard
FHIR (Fast Healthcare Interoperability Resources) unifies the format of record data such as cases so that it can be accessed by artificial intelligence systems and used for deep learning analysis. However, it still takes a lot of manpower and time to integrate existing medical data.
Other challenges include how to promote it to the clinic and establish the trust of physicians and patients on the results of the analysis data, and how to implement artificial intelligence system analysis and application in the medical workflow.
.