Currently, Syntiant's Neural Decision Processor (DNP) architecture is mainly used in speech recognition. Now, students at Harvey Mudd College (Claremont, California), in collaboration with Syntiant engineers, have shown that the AI processors are also suitable for other applications where low power consumption is important.
The system developed by the students is able to detect significant events from the constantly newly acquired sensor data using the NPD101. For this purpose, they collected and trained more than 60,000 movements and gestures of the hand on the basis of the development and deep-leaning tranig board "NDP9101" - from the motionless hand to reading a watch to hand rotations. As a result, the neural network achieved an accuracy of 94 percent.
"The students were able to gain experience in how machine learning, sensors and low power ICs work together and were able to develop a system that Syntiant is already introducing to customers," said David Harris, Professor of Engineering Design at Harvey Mudd College, who served as project consultant. They also showed how versatile the NDP architecture can be used.
"We will continue to focus on developing systems for "always-on" voice control, but we also believe this approach is well suited to enable portable devices to detect motion and gestures with very high accuracy and at much lower power consumption than traditional controllers," said David Garret, Syntiant's Vice President of Hardware Engineering.
Syntiant has developed the architecture of its Neural Decision Processors specifically to perform AI tasks based on a neural network. This is 100 times more efficient than with conventional controllers, and the throughput is increased tenfold. With this, Syntiant wants to bring AI into a variety of edge applications.