• Innovation through disruptive and scalable technology .
  • Cutting-edge AI research .
  • Accelerating innovations in research and service .
  • We strive for (and achieve) excellence! .
  • “SotA” (State-of-the-Art) .
  • Visual demo of research and service innovation
  • Human. Machine. Experience Together .

Meta AI

AI makes AI

Background

Meta AI refers to an AI system that can automatically learn from given data or adapt to new environments rapidly with minimal supervision of human experts.

In different fields, deep learning gained a great success and based on this, multifold application has been generated including software/hardware robots, drones, and self-driving cars. However, many experts should spend a lot of time and effort to successfully implement and train deep learning models. As solutions to reduce these costs, there are numerous on-going researches for the automation of most parts of deep learning. In this way, users can easily analyze the data they possess and furthermore, implement high quality deep learning system, even far better than the ones designed by the experts.

Project

In Meta AI system development, we should consider three main parts: creation of Meta AI interface, development of core algorithms and establishment of cluster infrastructure.
First, we create an interface for providing accurate information on the process of the system to users as well as receiving data and other inputs from them. We also design interfaces for the test of generated models and their distribution.

Second, we implement two kinds of core algorithms to execute the automation. One is the technology which automatically find the most appropriate neural network structure for learning given datasets. Previously, it took a long time due to the huge search space of network architectures, but recently, an architecture representation in the continual space is proposed, and this allows efficient search of the network architecture by optimization algorithms. Besides, we also try to implement a technique which can search not just predefined modules in the networks, but the entire structure of neural networks. The other is the automatic selection of hyperparameters used in the machine learning. For example, before the training, experts should properly set the hyperparameters – parameters required for the deep learning - based on their experience and knowledge by considering characteristics of assigned tasks, but this requires considerable amount of human resources as well as time. Through the automatic selection of the hyperparameters, we try to design services to provide models that can efficiently function following the users’ request and at the same time, minimize the involvement of experts as well as users.

Finally, to implement above mentioned algorithms, we establish GPU cluster infrastructure for the training of models and parallelizing search process, and design monitoring system to efficiently manage the entire system.

Furthermore, our research includes automation techniques in the machine learning, such as architecture optimization and usage of training data. For example, it is common to randomly assign weights in the neural network before the training, but to develop neural networks which can adopt faster to the new dataset, we research on meta-learning to find good initialization of the weights as well as transfer-learning to utilize existing weights of previous tasks.

Conclusion

Since Meta AI can serve as a great automatic tool, it is possible to utilize our Meta AI system across diverse projects related to deep learning. Moreover, in this hyper-connected and super-intelligent society, the demand for data processing keeps increasing; with this momentum, Meta AI will certainly provide benefit to numerous users in the nearest future.

References


  • Paul
  • Logan
  • Alex
  • Elliot
  • Dore
  • Sunny
  • Finn
  • YS
  • Muzi
  • Lily