Researchers created a nanosecond-scale volatile modulation system combining a phase-change material as a breakthrough for optical computing. The requirement for processing capacity has risen due to developments in computer vision and autonomous driving. The power and size constraints on current optical computer circuits have prompted the development of nonvolatile integrated photonics. Phase-change materials (PCMs) have the potential to be used in neuromorphic photonic processors and photonic memory. However, there are difficulties with online training, such as frequent and quick switching, making it essential to get beyond these problems for effective and speedy training.
A 5-bit photonic memory with quick volatile modulation and a nonvolatile network for quick training was created by researchers. The memory preserves recorded weight information using a PIN diode carrier dispersion effect. It also employs the diode as a microheater for multiple reversible phase shifts, producing an energy-efficient photonic computing technique.
The study team modeled an optical convolutional kernel architecture using the shown photonic memory and its operating concept. Surprisingly, they recognized the MNIST dataset with a recognition rate of above 95%. This demonstrates the viability of quick training using volatile modulation and weight storage using nonvolatile modulation of five bits.
This ground-breaking study opens the door for a brand-new paradigm in photonic memory. It provides a viable option for using nonvolatile components in fast-training optical neural networks. With these developments, optical computing has a more promising future than before.
Related Content: Leaky-Wave Metasurfaces – Free Space, Integrated Systems