Light field microscopy (LFM) offers rapid volumetric imaging of dynamic biological processes via 2D light field snapshots. However, LFM's 2D-to-3D inversion with limited spatial-angular tomography leads to artifacts and poor axial resolution. Here, we introduce light-field meta neural representation (LFMNR), a new LFM reconstruction paradigm based on physics-informed implicit neural representation and meta learning, to address these issues for LFM. Leveraging INR's continuity, LFMNR achieves self-supervised, artifact-free 3D reconstruction with enhanced structural fidelity (~2.2-fold improvement), spatial resolution (~4.4-fold enhancement) and data compression (~10-fold), when compared to classical model-based light-field deconvolution. In addition, our meta-learning and progressive sampling strategies also mitigate INR's intrinsic limitations in weak generalization and low representation speed scene by scene, thereby resulting in rapid representation (~100-fold acceleration) of hundreds of consecutive volumes capturing sustained biological dynamics in three dimensions. LFMNR demonstrates superior performance across diverse biological samples without any prior spatial structure knowledge. We showcase LFMNR's capability in observing cell apoptosis in several hours and capturing instantaneous organelle interactions at millisecond timescale. LFMNR approach readily delivers high-fidelity, high-speed 3D imaging with vast potential applications in biological research and provides a paradigm shift with extensive value for computational imaging reconstruction.

High-fidelity, high-resolution light-field reconstruction with physics-informed neural representation learning