# Diederik Kingma > ~1984– · Computer Scientist, AI Researcher > > **Recorded contribution:** Adam optimizer (used in virtually all modern neural network training); VAE (variational autoencoders) ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation Dutch machine-learning researcher Diederik P. Kingma co-developed variational autoencoders with Max Welling and the Adam optimizer with Jimmy Ba. The 2013 VAE paper introduced a practical reparameterized variational objective for continuous latent-variable models; the 2014 Adam paper combined adaptive second-moment scaling with momentum-like first moments. Both became widely used, though neither is universally optimal. Kingma's variational autoencoder work made latent-variable learning practical by rewriting stochastic inference so gradients could pass through sampled variables. Adam addressed a different problem—how to scale each parameter's update from estimates of its recent first and second gradient moments. ## 2. The problem inherited Latent-variable neural models needed low-variance gradients through stochastic variables, while deep-network optimization needed useful parameter-wise step scaling under noisy minibatch gradients. ## 3. The central contribution Kingma co-created the VAE's reparameterized learning method and Adam's adaptive stochastic optimizer. ## 4. Reconstruct the mechanism 1. Encode an observation into parameters of an approximate latent posterior. 2. Sample latent noise and transform it differentiably by the posterior parameters. 3. Optimize reconstruction likelihood plus a divergence regularizer—the evidence lower bound. 4. For Adam, maintain bias-corrected first and second gradient moments and scale each parameter update accordingly. ## 5. What changed downstream - VAEs became foundational generative and representation-learning models. - Adam became a default optimizer across deep learning because it often trains effectively with limited tuning. - The reparameterization trick became a standard tool for amortized variational inference, while Adam became a default optimizer across deep-learning experiments and lowered the friction of training large models. ## 6. Attribution, limits, and uncertainty - Both are two-author results with extensive probabilistic-inference and optimization predecessors. - VAEs may trade sample sharpness for tractable learning; Adam can generalize or converge worse than alternatives and its ubiquity is not proof of universal superiority. - A VAE's evidence lower bound trades reconstruction against an approximate posterior and can yield unused latents; Adam's convenience can hide sensitivity to learning rate, weight decay, numerical details, and generalization. ## 7. Reconstruction lab Train a two-dimensional VAE on synthetic clusters and plot its latent space. Then optimize one function with SGD and Adam using equal budgets and explain the different trajectories. Plot reconstruction loss and KL divergence separately, then train the same model with Adam and SGD so optimizer convenience is not confused with a better probabilistic model. ## 8. Evidence trail - [Auto-Encoding Variational Bayes](https://arxiv.org/abs/1312.6114) — arXiv - [Adam: A Method for Stochastic Optimization](https://arxiv.org/abs/1412.6980) — ICLR / arXiv - [Diederik P. Kingma publications](https://scholar.google.com/citations?user=yyIoQu4AAAAJ) — Google Scholar --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*