How Actapio Automates AI Models Using GPUs

By automating its machine learning processes using NVIDIA GPUs, Actapio, Inc., a subsidiary of Yahoo! Japan Corp., is helping its customers get to market faster.
by Ahana Dave

Traditional machine learning and data science project cycles are time-consuming and repetitive. From data cleaning to model validation, almost every stage involves cumbersome, manual processes.

By automating its machine learning processes using NVIDIA GPUs, Actapio, Inc., a subsidiary of Yahoo! Japan Corp., is helping its customers get to market faster.

Based in Washington state, Actapio offers consulting and computing services to help U.S.-based startups enter the Japanese market. Its customers offer a vast range of consumer applications, from advertising and ecommerce to speech recognition and data services.

To speed its processes, Actapio’s AI research and development team has developed Hilbert, a framework that creates production-level AI models to bring deep learning capabilities to over 100 consumer services powered by Yahoo! Japan.

From email to video streaming, the framework improves recommender systems, click-through rate prediction and search query analysis.

Hilbert also shortens turnaround times for trial and feedback when developing AI models. For example, model structure, input feature, data transformation and hyperparameter tuning settings are all defined within Hilbert’s configuration files. This allows know-how to be easily shared between services by simply copying and pasting config files.

In addition, the knowledge gained from successful services and the new AI models developed by each service are incorporated into the framework. As a result, developers can take the good results from one service and immediately try them out on another. This helps data scientists and engineers more quickly understand the efficiency of these smart models. 

Browsing Challenges and Opportunities

Not all of Actapio’s data science teams have enough scientists, yet these applications required cutting-edge AI technology. That’s why the company needed a framework that would automatically create efficient and working AI models. They also faced challenges in figuring out how to optimize AI models efficiently.

Using nearly 300 NVIDIA V100 GPUs at its new hydroelectric-based data center helped Actapio reduce model training time and optimize its performance daily. The team saw performance improvement of up to 65 percent across various Yahoo! services, including shopping, ebooks and travel.

Using the Hilbert framework, the team fully automated model optimization processes, such as model architecture search and hyperparameter tuning, and introduced them into actual services. This allows for day-to-day model optimization using the latest data.

“In our experience, the fresher the data, the better the performance,” said Norifumi Matsuya, CEO of Actapio. “NVIDIA products offer high performance and stability, so we can use them with confidence.”

Looking ahead, AI, machine learning and deep learning are becoming integral parts of Actapio’s services. The company believes there’s still a lot of room for innovation in the area of recommendations and click-through rate prediction. It aims to further close the go-to-market gap between image, speech and translation with recommender and click-through rates to increase the relevancy of products that customers are searching for.