KAIST Cuts AI Fine-Tuning from 10 Hours to 3… Faster Development of Medical and Coding AI (IMAGE)
Caption
Figure 1. Conceptual diagram of the clustering-based balanced data sampling and allocation method developed by the research team. Different colors represent clusters of data that have similar effects on learning, and different shapes within the same color represent specific data types within each cluster. In conventional training on the full dataset, data types with a large number of samples can be repeatedly included across multiple GPUs, so similar computations may be repeated. The proposed method selects diverse data from each cluster in a balanced way and allocates different clusters to each GPU, which reduces redundant computation. For
Credit
KAIST
Usage Restrictions
No
License
Licensed content