Few-shot transfer learning using a recurrent neural network for hologram reconstruction (Conference Presentation)

Abstract

We report a novel few-shot transfer learning scheme based on a convolutional recurrent neural network architecture, which was used for holographic image reconstruction. Without sacrificing the hologram reconstruction accuracy and quality, this few-shot transfer learning scheme effectively reduced the number of trainable parameters during the transfer learning process by ~90% and improved the convergence speed by 2.5-fold over baseline models. This method can be applied to other deep learning-based computational microscopy and holographic imaging tasks, and facilitates the transfer learning of models to new types of samples with minimal training time and data.