A UCLA research team led by Aydogan Ozcan has been developing a diffractive deep neural network, a machine that combines optical diffraction deep learning with light-matter interaction. The goal for the team was to engineer diffractive surfaces that collectively perform optical computations at the speed of light. Prototype of a broadband diffrative neural network. Courtesy of: Ozcan Research Group, UCLA. Ozcan, a professor of electrical and computer engineering at UCLA, noted that the team’s previous diffractive network models were developed to process information through a single wavelength channel, therefore requiring a monochromatic and coherent illumination source. “The new results that we have published this week in Light: Science & Applications address this limitation and teach us a method to process information at a continuum of wavelengths,” Ozcan said. “To prove the efficacy of this broadband diffractive network framework, we have designed optical networks that can process broadband optical pulses that are input to a given network.” Since the connection between the input and output planes of a diffractive neural network is established via diffraction of light through passive layers, the inference process and the associated optical computation does not consume any power except the light used to illuminate the object of interest. “Diffractive optical networks that we have created provide a low-power, low-latency and highly scalable machine learning platform, with various potential applications in robotics, autonomous vehicles, and defense industry, among many others,” Ozcan said. Ozcan’s team demonstrated the success of their framework by designing and implementing various optical components, such as single and double passband spectral filters and a spatially controlled wavelength de-multiplexing system using a broadband THz pulse as input. The latter’s input light pulse is filtered by a trained diffractive network into four different bands at four different locations at the output plane. “Simultaneously analyzing and processing light across many wavelengths presents unique opportunities to enhance the inference and generalization capabilities of diffractive optical networks to perform machine learning tasks,” Ozcan said. Ozcan believes these opportunities for enhancement could manifest in all-optical object recognition for robotics and autonomous cars and designing deterministic and task-specific optical components, such as ultra-thin lenses or smart camera systems, expanding the optical design space beyond human intuition.