Short Floating-Point CNN Accelerator for Brain-Computer Interface to Decode Visual Information.

In the context of a brain-computer interface (BCI) system, a prior study proposed a method to reconstruct original viewed images from visually evoked brain activity by using an AlexNet-based generative adversarial network. In this paper, we present an accelerator using a short-bit-length data format that can process this BCI decoding. We implemented the decoding process on the proposed accelerator with a 10-bit floating format consisting of the sign, a 6-bit exponent, and a 3-bit mantissa for 256x256 RGB video image generation. By introducing the 3-bit mantissa for the 23-layer operation, including convolution, deconvolution, and leaky ReLU, the image was decoded at a peak signal-to-noise ratio of around 30 dB. The implementation of the proposed accelerator in a single field-programmable gate array exhibited performance of 168 GOPS at a clock speed of 250 MHz.