https://quietsignalslab.com/writing/ Long-form technical writing by Alexander Hepburn on ML privacy, membership inference, retrieval systems, and the engineering problems behind privacy-sensitive AI. Four to eight pieces per year. en-gb © 2026 Alexander Hepburn hello@quietsignalslab.com (Alexander Hepburn) hello@quietsignalslab.com (Alexander Hepburn) Thu, 01 May 2026 00:00:00 +0000 https://quietsignalslab.com/writing/multi-task-learning-distribution-shift/ A model trained on one domain and deployed in another will lose performance. This piece works through four controlled ablations on cross-domain NLP classification — cross-domain training, data duplication, label granularity, and cross-domain evaluation — to isolate when multi-task learning acts as a structural buffer against distribution shift, and when it doesn’t. Mon, 14 Apr 2026 00:00:00 +0000 https://quietsignalslab.com/writing/multi-task-learning-distribution-shift/ https://quietsignalslab.com/writing// https://quietsignalslab.com/writing// ========================================================================= -->