KAIST Develops AI Chip That Recognizes Changes in Motion over Time (IMAGE)
Caption
Figure 4. AI computing based on temporal information using a 3D multi-timescale reservoir
Devices with different ionic dynamics were stacked vertically to create a dual-timescale reservoir that processes fast and slow changes over time simultaneously. The reservoir distinguished 16 different reservoir states from 4-bit temporal patterns. In the classification of temporal information using Moving-MNIST, a dataset of moving handwritten digits, it processed inputs on cycle scales of 1 ms and 10 ms and achieved validation accuracies above 90% at both timescales.
Credit
KAIST
Usage Restrictions
No
License
Licensed content