Researchers Present An Optical Chip That Can Train Deep Neural Networks Using Direct Feedback Alignment

McKinsey has recently reported that Machine Learning applications have seen a skyrocketing rise of $165 billion yearly. But any Machine Learning model must be trained before performing any kind of task.

But training is not an easy task. The training of Tesla’s Artificial Intelligence system might cost several million dollars in terms of its electric power consumption and requirements like supercomputer kind of infrastructure. A team consisting of researchers from George Washington University, Queens University, University of British Columbia, and Princeton University collaborated on a paper where they presented an optical chip that can be used for training the machine learning hardware. 

A widening gap has been left in between the demand for AI and computer hardware because of the rapid evolution of AI. A possible solution to computer hardware problems is simple optical chips or photonic integrated circuits as they have a viable solution to develop higher computing performance measured in terms of the number of operations per second per watt consumed (TOPS/W). However, it has been shown that these devices can improve the efficiency of machine intelligence used for data classification. But these photonic chips are to be enhanced in the machine training and front-end learning schemas.

Machine Learning, in general, is a two-step process where in the first step, the data is used to train the system, and in the second step, the performance of the AI system is tested using the test data. The research team has observed an error after a single training step and reconfigured the hardware systems for a second training process, and additional training cycles have also been performed until an acceptable performance is reached. Until day photonic chips have only demonstrated their ability for data inference and classification, but researchers have also demonstrated that it is also possible to speed up the training process too. This enhanced AI capability is part of a larger effort centered on photonic tensor cores and other electronic-photonic application-specific integrated circuits (ASIC) that take advantage of photonic chip manufacturing for machine learning and AI applications.

These advancements are much needed in the AI, and semiconductor industry as this hardware will help speed up the training of the machine learning process. Also, training these AI systems can cost a huge amount of energy and carbon footprint, and training using these photonic chips can help us reduce these expenses.

Check out the Paper and Reference Article. All Credit For This Research Goes To Researchers on This Project. Also, don’t forget to join our Reddit page and discord channel, where we share the latest AI research news, cool AI projects, and more.