Kakao Holds Second Internal Tech Seminar Techtalk, Sharing TF Serving 10x Optimization

Kakao held the second session of its internal technical seminar Techtalk, sharing practical experiences from developers working on AI and machine learning. This seminar was conducted for Crews interested in recommendation technology and deep learning, focusing on the learning and execution process.

TensorFlow Serving 10x Optimization

Ha Kwang-sung, an engineer from the Kakao Recommendation Team, introduced the process of achieving 10x faster speed compared to TensorFlow Serving when using TensorFlow models in services. The initial prototype had performance similar to existing serving, but performance was improved by 2x through layer fusion and code-level optimization, and by more than 5x through domain-level optimization. Subsequently, bottleneck areas were intensively optimized, resulting in a final speed improvement of more than 10x.

Large-Scale Data Training and Direction for Recommendation Systems

Engineer Han Min-ho shared methods for maintaining model performance while processing 300GB of data. He covered a case study from the Ad Recommendation Team, where Hadoop bottleneck analysis and data reduction techniques were applied to solve data volume and processing time issues while developing a deep learning conversion optimization targeting model.

Analyst Choi Gyu-min discussed the direction and goals for personalized recommendation systems, referencing Netflix's recommendation system. The presenters evaluated that the process of organizing knowledge and receiving feedback through seminars contributes to personal growth and the formation of internal culture. The next Techtalk is scheduled for May.