Durk Kingma

Durk Kingma is a computer scientist whose work includes the variational autoencoder (VAE) with Max Welling and the Adam optimizer with Jimmy Ba. These collaborations received ICLR Test of Time awards in 2024 and 2025, respectively. He publishes as Diederik P. Kingma; Durk is his Frisian nickname.[2][3][4][7][8]

Kingma led research projects on generative models for text, image, and video as a research scientist at Google Brain / DeepMind from 2018 to 2024. His personal website, as of September 18, 2026, lists research on large-scale machine learning at Anthropic beginning in 2024. His earlier doctoral work provides context for the probabilistic inference methods behind the VAE.[2][3]

Kingma worked as a junior research scientist in Yann LeCun’s lab at New York University in 2009 and 2012. Between those periods, he co-founded Advanza and served as its technical lead from 2010 to 2012; the company was acquired in 2016. From 2013 to 2017, he pursued a PhD at the University of Amsterdam under Max Welling, studying deep learning and generative models. He graduated cum laude with the thesis Variational Inference and Deep Learning: A New Synthesis, and spent the summers of 2014 and 2015 collaborating at DeepMind.[2]

Kingma and Welling initially submitted Auto-Encoding Variational Bayes in December 2013. Their method addresses learning and inference in probabilistic models with continuous latent variables, including cases where the posterior distribution is intractable. Under mild differentiability conditions, reparameterizing the variational lower bound yields an estimator that can be optimized with standard stochastic gradient methods. For independent, identically distributed data with continuous latent variables per datapoint, they also fit an approximate inference—or recognition—model to the posterior using that estimator.[3]

The joint paper won the inaugural ICLR Test of Time award in 2024. The award citation credited it with integrating deep learning and scalable probabilistic inference, giving rise to the VAE. It also acknowledged concurrent work by Rezende and colleagues proposing a similar idea, published at ICML 2014.[7]

With Jimmy Ba, Kingma introduced Adam: A Method for Stochastic Optimization, initially submitted in December 2014 and published at ICLR 2015. Adam adjusts parameter-specific learning rates using estimates of the first and second moments of gradients. The authors described it as suited to large problems, noisy or sparse gradients, and non-stationary objectives. Their paper also introduced AdaMax, a variant based on the infinity norm.[4][8]

Kingma joined OpenAI’s founding team in 2015 and worked there as a research scientist until 2018. He led its Algorithms team, focused on basic research. A decade after Adam’s conference publication, the paper won the 2025 ICLR Test of Time award; the award citation described Adam as one of deep learning’s most widely adopted optimization algorithms.[2][8]

Kingma and Prafulla Dhariwal proposed Glow in 2018, a generative flow using invertible 1×1 convolutions. Flow-based models offer tractable exact log-likelihood and exact latent-variable inference, with parallelizable training and synthesis. The authors reported that optimizing Glow with a plain log-likelihood objective enabled realistic-looking synthesis and manipulation of large images.[5]

In Variational Diffusion Models, published at NeurIPS 2021, Kingma worked with Tim Salimans, Ben Poole, and Jonathan Ho on likelihood-based diffusion modeling. They expressed the variational lower bound through the signal-to-noise ratio of diffused data and showed that its continuous-time form is invariant to the noise schedule except for the endpoint signal-to-noise ratios. This result let them learn a noise schedule that minimizes the bound estimator’s variance; they reported faster optimization. They also used the model in a bits-back compression scheme and reported lossless compression rates close to the theoretical optimum.[6]