A hitchhiker's guide to Poisson gradient estimation

Researchers have proposed new methods to improve Poisson gradient estimation and decision-focused learning in machine learning models.

Poisson-distributed latent variable models are widely used in computational neuroscience, but differentiating through discrete stochastic samples remains challenging^[1]. Two approaches address this challenge: Exponential Arrival Time (EAT) simulation and Gumbel-SoftMax (GSM) relaxation. A modified EAT method has been shown to reduce second-moment bias and exhibit better overall performance than GSM on variational autoencoders with Poisson latents^[1]. Meanwhile, researchers have also made advancements in decision-focused learning (DFL), a paradigm that trains to directly minimize a task loss. A proposed method combining stochastic smoothing with score function gradient estimation works on any task loss and opens up the use of DFL methods to nonlinear objectives and uncertain parameters^[2]. Additionally, a novel unbiased quantization routine called MS-EDEN achieves more than 2x lower quantization error than stochastic rounding in NVFP4 pre-training, and is supported in hardware by NVIDIA Blackwell GPUs^[3].