Energy-Efficient Near-Pixel 2D CNN Accelerator

1.1 Data Processing System (Novel Computing, Machine Learning, Neural Network, etc.)
1.2 Data Storage System (Computation-in-Memory, Persistent Memory, etc.)
1.4 Data Generation System (Sensor, Microsensor, etc.)

Makoto Takamiya

Institute of Industrial Science

Professor

The convolution operation, an original feature of the convolutional neural network (CNN) algorithm, is performed only near the featured pixel. The CNN operation is two-dimensionally performed using the digital integrated circuit within the vicinity of the pixel, without writing data to the external memory.