RIKEN Develops a Method to Apply CNN to Non-Image Data
RIKEN found that MATLAB has useful features for deep learning and machine learning that enable users to speed up research by reducing coding time.
Key Outcomes
- Apply CNN to non-image data for quick feature extraction and classification
- Reduce coding time
- Use multiple GPUs for deep learning training and inferences
Detecting small variations in phenotypes or class labels, such as those used in genome analysis, is an important but challenging task. Even with sufficient data, identifying genes or relevant features is not easy.
It is generally known that CNN is an effective method for image data, but it was harder to apply for genetic research that requires handling of non-image data such as RNA sequence data.
RIKEN is the largest research organization for basic and applied science in Japan. In their DeepInsight project, the researchers converted non-image data to image format to apply CNN effectively. Using MATLAB® to place similar elements together in a cluster, the team performed feature extraction of non-image data and identified hidden mechanisms.
