Transformers are Bayesian Networks — mentioned by asgeralstrup

Fri, 27 Mar 2026 11:06:03 GMT

Built a project that goes from CSV or BIF to transformer using Claude Code. It can learn a Bayesian network from data, compile it into transformer weights so the forward pass performs belief propagation, and export the result to ONNX. The fastest batched ONNX path hits 3 to 6M rows/sec with OpenVINO on consumer GPU. The BN transformer handles missing data naturally. Since inference is just belief propagation, you can omit any feature from the evidence and the model marginalizes over it automatically — no imputation needed. In benchmarks, it retains 90% of its accuracy when 50% of test…

Paper and discussion history. A mention is not an endorsement.

Transformers are Bayesian Networks — HN submission

Fri, 20 Mar 2026 20:50:27 GMT

Transformers Are Bayesian Networks

40 points; 33 discussion replies on HN. Paper and source history. A submission is not an endorsement.

Transformers are Bayesian Networks · HackerPapers discussions

Transformers are Bayesian Networks — mentioned by asgeralstrup

Transformers are Bayesian Networks — HN submission