Machine learning at 10,000 times faster than CPU…

The University of Michigan research team announced that it has developed a computing device that can perform machine learning at 10,000 times the speed of the CPU currently installed in computers through a resistor called a memristor.

Memristors are passive devices with properties that can store electric charges and change resistance. The memristor serves as a logical operation unit and a memory element. As a memory device, non-volatile memory can be used to hold analog data, and redundant computation (FMAD) is also possible.

The advantage of adopting a memristor as a processing device is that the memristor can serve as both a memory and a computing device, so it can perform calculations suitable for neural networks. The research team argues that GPUs are 10 to 100 times better than CPUs in terms of power consumption and throughput, but Memristor has the potential to be 10 to 100 times better than GPUs.

The research team actually created a prototype model of a memory star chip that converts between 5,800 memristors, an OpenRISC CPU, a communication circuit, and between analog and digital, and created a program that performs machine learning algorithms.

As a result of actual machine learning using this chip, it is said to have achieved 100% accuracy in the basic perceptron that discriminates Greek characters and the Sparse Modeling that identifies and optimizes the classification of image patterns. In the dual neural network machine learning, which finds commonalities and differences in breast cancer test data, it was possible to classify malignant and benign cancers with 94.6% accuracy.

Of course, the research team explained that the analog information maintained inside the memristor has a problem with reliability and that there are challenges to be used commercially. He said he will continue research to solve this problem. Related information can be found here .