The so-called “digital healthcare transition” is the core of the development of a multilateral connection system based on efficiency, economic efficiency, convenience, speed, and future development in a one-on-one relationship with simple service provision. This is because there are no high-performance algorithms in the world for all problems.The same is true of deep learning technology, which has recently been widely used in the field of artificial intelligence. Various structures must be applied to find the deep learning model that performs best for a given problem, especially in order to find the optimal model, various hyperparameters unfamiliar to non-experts such as learning rate, batch size, and optimization method are applied.Superparameters refer to parameters that users (research developers) must specify in advance before starting learning artificial intelligence models through data. The number of algorithms to be tested can vary from hundreds to tens of millions, and it takes a long time to develop. Since the computing devices given to artificial intelligence developers are limited and the development period is not infinite, trying several combinations based on expert intuition or so-called know-how is a common process to develop artificial intelligence models.Although artificial intelligence provides automated convenience to human life, it would have required excessive effort by numerous experts as described above before practical artificial intelligence products were developed. Given the limited resources and time, global conglomerates are striving to secure AI experts because skilled experts are needed to develop AI technology faster than competitors, raising the salaries of experts in this field. Currently, AutoML mainly deals with feature engineering automation, hyperparameter auto-optimization, pipeline auto-optimization, and neural architecture search (NAS). Feature engineering mainly refers to a method of creating new input functions that can increase the performance of artificial intelligence models using input variables. Feature engineering has also been mainly done by the insights and experiences of experts in the field, and various methods are being attempted to automate it. However, most technologies have the disadvantage of requiring a very large amount of computation because they try a wide variety of functional conversion methods and choose the best method among them.Suparameter optimization is an essential task for artificial intelligence developers as described above, and because no one knows the answer, it is a tedious process that must be found continuously repeating the Trial-and-Error process, and waiting for results with a prayer. The performance change may be very large depending on the hyperparameter setting, so it is a very important process in the development of the entire artificial intelligence model. Therefore, some compare this process to “finding a needle on a sandy beach,” and others to “art.”Therefore, it was expected that hyperparameter automatic optimization could help artificial intelligence developers and researchers make their lives a little more relaxed. The most widely used methods are methods based on Bayesian Optimization. Simply put, based on the performance of previously used superparameter combinations, the performance of previously unused superparameter combinations is predicted in advance, and some of the best combinations are selected and automatically applied to the next experiment.
This process is repeatedly performed until a satisfactory result is obtained.It shows faster and better performance than grid search or random search, which attempts all combinations, but this also has a limitation in that as the size of the data increases, the amount of computation will inevitably increase exponentially. In particular, if you want to apply it to healthcare big data including various types of heterogeneous data, more careful attention may be needed.Feature Engineering described above, hyperparameter optimization, as well as Pipeline automatic optimization, which automatically optimizes a series of entire processes required for artificial intelligence development, are also being introduced. Software, software toolkit, and software services, including these functions, are available for free or for a fee. Since Auto-WEKA, Auto-sklearn, and TPOT are the most widely known free AutoML software or software toolkit, artificial intelligence research and developers can easily use it.Recently, cloud computing services have been introduced around global IT giants, and in particular, Google, Microsoft, IBM, and Amazon are distributing them, including AutoML functions, most of which advocate automatic Pipeline optimization. These Pipeline auto-optimization software provide services that find the optimal model in their own efficient ways (related papers and promotional materials introduce a number of cases that match or exceed the performance of human experts directly optimizing), but this also cannot be considered to work well for all problems and the amount of computation cannot be ignored. In particular, if you apply the AutoML function, which uses cloud computing to charge, you may have an absurd experience where you can easily be charged for a PC even though you haven’t used it for a long.Deep learning technology has recently achieved the most innovations in the field of artificial intelligence. Deep learning is an extended version of the Artificial Neural Network (Artificial Neural Network) algorithm, which can be thought of as an implementation of an artificial neural network structure with a deeper and more complex structure than the previously widely used structure. Researchers have been learning by setting up neural network structures in advance, and subsequent researchers have often used structures verified by many people as they are.However, recently, research in the field of automatically exploring even this neural network structure based on data has attracted a lot of attention, which is the field of neural network structure search (NAS). Domestic and foreign researchers are showing excellent results on various tasks using NAS, and the structure of the finally explored neural network can be more complex and bizarre for interpretation or intuitive analysis. Therefore, when interpreting results is applied to important healthcare issues, these shortcomings must be considered.”Then, despite these difficulties, there are companies that go beyond commercialization and commercialization to be verified (medical devices) for certification permission to certify the Ministry of Food and Drug Safety.Lee Jae-yong, CEO of “Info Mining,” is the mastermind. In fact, it has already developed its own platform called ConnectDac to measure and manage the health information of wearable watch-type devices to individuals. There are already many companies in devices and platforms, so what sets them apart from other companies? It represents the measurement closest to medical. It advocates true care and management, not a simple approach to healthcare.U.S. hospitals and Korean university hospitals are already receiving licenses and development costs for platforms that have been verified for use. (TELADOC HEALTH CARE, YEONSEO UNIV. Severance Hospital…)”
‘
▲ INFOMINING CO., LTD.
▲ CEO : Jae-yong Lee
▲ http://infomining.co.kr
▲ worldconsult@infomining.co.kr
▲ +82-70-4914-2970
EJ SONG
US ASIA JOURNAL
