About

Binarized neural networks

Binarized neural networks (BNNs) use only 1 bit to encode each parameter, instead of 32 bits. This dramatically reduces the memory consumption: less accesses from a smaller memory results in a theoretical speed up far larger than 32 times. Plumerai’s proprietary architecture is implemented on off-the- shelf available reconfigurable chips (FPGAs) to perform inference of BNNs efficiently. Globally there will be 100 million autonomous drones and robots, initially mostly used for delivery. FPGAs can be easily integrated in these products and are an ideal platform for bringing BNNs to market.