Machine-learning model generates images using light

As machine-learning models become more powerful and complex, it is increasingly challenging to implement them efficiently on conventional computers. This has created interest in building physical systems with intrinsic behaviour that can be used to implement those machine-learning models directly. Instead of performing computational operations on a digital computer, these systems function according to their own physical laws, with the goal of reducing the time and energy requirements of computation. Writing in Nature, Chen et al.¹ report a model that uses a combination of digital and optical components to generate new (unseen) images. The system directly modifies a beam of light to create pictures of numbers, objects and van Gogh-style paintings.