# David Blei > ~1977– · Computer Scientist, ML Researcher > > **Recorded contribution:** Latent Dirichlet Allocation (LDA); topic models; variational inference ## 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 American computer scientist David Blei co-developed latent Dirichlet allocation (LDA) with Andrew Ng and Michael Jordan and became a leading researcher in probabilistic topic models and variational inference. His work asks how hidden structure can explain observed collections—documents, images, scientific data—while making uncertainty explicit. LDA's 2003 publication turned topic discovery into a reusable generative-modeling problem. ## 2. The problem inherited Large document collections contained recurring themes, but manual taxonomies were expensive and simple word counts did not explain how multiple latent topics combine inside one document. ## 3. The central contribution Blei co-created LDA, a probabilistic generative model in which documents mix latent topics and topics are distributions over words, and advanced scalable approximate inference for such models. ## 4. Reconstruct the mechanism 1. Draw a topic-mixture distribution for each document. 2. For each word position, sample a latent topic from that mixture. 3. Sample the observed word from the chosen topic's word distribution. 4. Invert the generative story with posterior inference to estimate topics and document mixtures from observed text. ## 5. What changed downstream - LDA became a standard entry point for unsupervised text exploration and probabilistic modeling. - Variational-inference research made complex Bayesian models practical on larger datasets. ## 6. Attribution, limits, and uncertainty - LDA is a three-author result grounded in earlier latent semantic and mixture modeling; discovered topics are not objective truths. - Bag-of-words assumptions discard order, topic number is chosen, and interpretation can be unstable or misleading across preprocessing and populations. ## 7. Reconstruction lab Create a six-document corpus about two overlapping themes. Fit or manually iterate a two-topic mixture, inspect high-probability words, then change one document and quantify topic instability. Run the model with one, two, and four topics and compare held-out likelihood with human interpretability. Permute random initialization to see whether apparently meaningful themes persist. Inspect a document whose mixture is poorly explained and resist renaming every statistical cluster as a real social category. LDA’s value is a transparent generative decomposition of co-occurrence; its latent variables are useful hypotheses, not discovered essences, and approximate inference adds another source of variation. Have two readers independently label the inferred topics and report disagreement, separating computational reproducibility from semantic agreement. ## 8. Evidence trail - [Latent Dirichlet Allocation](https://www.jmlr.org/papers/v3/blei03a.html) — Journal of Machine Learning Research - [David Blei](https://www.cs.columbia.edu/~blei/) — Columbia University - [Variational Inference: A Review for Statisticians](https://arxiv.org/abs/1601.00670) — Journal of the American Statistical Association --- *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.*