Fig. 1 Deep learning-based image recognition and the new SSD
[Click to enlarge image]

A Japanese research group developed what it claims is an SSD (solid state drive) best suited for deep learning-based image recognition.

The group is led by Ken Takeuchi, professor at the Faculty of Science and Engineering, Chuo University. The new SSD, "Value-Aware SSD," evaluates the value of image data and is controlled so that important and not-so-important data are stored in high- and low-reliability memory cells, respectively.

At data centers, etc, SSDs are used as compact, high-speed, low-power-consumption storage devices. In the future, for real-time stream processing, etc, an enormous amount of data will be used for learning and deduction by using deep learning, creating a need to use SSDs for deep learning.

For the new SSD developed for deep learning, the research group developed three memory control technologies and succeeded in realizing a high reliability, high speed and high image recognition accuracy that enable to withstand deep learning applications.

The figure 1 shows the details of deep learning-based image recognition and the newly-developed SSD. For image recognition, for example, the characteristic vectors of an input image (taken by a camera, etc) are compared with the characteristic vectors of an enormous number of human images registered in the SSD so that the person in the input image can be identified.

1. Value-Aware Data Mapping (VADM)
One of the memory control technologies that the research group developed this time is "Value-Aware Data Mapping (VADM)." In this research, the group exploited the fact that data have their value. For example, in the case of 32-bit characteristic vectors, high-order bits have higher values than low-order bits.

Go to next page