Convolutional neural networks (CNN) are widely used in the latest 2D data processing methods—and, especially, image recognition and classification. List recently developed a calculator that makes it possible to execute CNNs optimally based on impulse models*, which, theoretically, can reduce the hardware resources and energy required to process data (the “inference” step).

The use of impulse models for CNNs requires coding the data as impulses that are transmitted between a wide range of artificial neurons arranged in different types of layers. To get the most out of the data coded in this way and complete the calculations as efficiently as possible, a special hardware architecture is also required. List developed its NeuroSpike calculator for this purpose. It is the first hardware architecture that enables the efficient calculation of a complete CNN’s impulse inference.

NeuroSpike uses eleven times less energy and is four times faster than impulse-based architectures used to calculate 2D convolution inferences. It is protected by three patents.

*Impulse models are a way of processing data coded as impulses transmitted within a neural network. The artificial impulse neurons use less complex and more energy efficient operators than traditional models, but they require more memory.