Digital optical neural network

MIT researchers came up with a novel artificial optical neural network that uses light to transmit activation and weighs data. The researchers called it a digital optical neural network. This system with only dew percent points of accuracy cost can achieve transmission energy and adding up 1000x over the traditional electronic devices. 

The research of the digital optical neural network is published under the title ‘Freely scalable and reconfigurable optical hardware for deep learning’ in the Nature’s Scientific Reports. The digital optical neural network handles the issues of power consumption in the neural network by replacing it with electric currents with optical signals. 

Digital optical neural network with its constant energy usage has empowered it to scale up for the other big deep learning models with its cost-effective and high performance rate. One 8 bit MAC operation only needs 3 femtojoules when compared to the thousand plus fj needed on an electronic chip. 

Most of the large deep learning models need large compute and memory resources for both inference and training. The Digital optical neural network also involves accelerator handwares such as TPs or GPUs to do computations quickly. But these hardwares can take up a lot of energy while it is running and so chip designers are continuously exploring new ways to make these machines more stronger and efficient by keeping memory close to computation elements. 

Optical data transmission is a cost-effective and more efficient option for long distance data transmissions.  But when coming to optical computing for digital signal processing, it is one of the active areas of research for decades. The progress of photonic integrated circuits is creating a huge interest in deep learning technology. And the linear algebra functions with optics that can be used by neural networks by applying light instead of electricity. 

But in the new approach, neural networks there is no need for analog computation and this new method can allow less noise and has more accuracy. The team of the researchers say that there are few errors that can be mitigated with corrections schemes while others only affect accuracy by less than 3%.