Blog & Notes
-
🛡️ AI Safety & Security: A Primer for the Agentic Era
Safety and security: two seventy-year-old problems that agents finally fuse. A primer on where AI safety meets AI security — and why that seam is the story of the next few years.
-
🧠 The Context We Can't Afford to Lose: Neural Networks from First Principles
The artificial neuron is from 1943. The learning rule, 1958. Backprop, 1986. A primer that builds the foundations of neural networks from first principles — and previews the lineage all the way to today's LLMs and agentic systems.
-
🌐 Isometric MT: Neural Machine Translation for Automatic Dubbing
Teaching a transformer to translate within ±10% of the source length directly — no N-best generation, no re-ranking — for length-matched automatic dubbing.
-
📄 Transfer Learning in Multilingual Neural Machine Translation with Dynamic Vocabulary
A shared dynamic vocabulary lets an NMT model grow to new language pairs by transferring parameters — up to +13.63 BLEU while training in a fraction of the steps.
-
📝 Paper Review: Phrase-Based & Neural Unsupervised Machine Translation
Work investigates how to learn to translate when having access to only monolingual corpora in each language.
-
✍️ Neural Machine Translation into Language Varieties
Training neural machine translation to produce specific language varieties and dialects, across pairs like Brazilian/European Portuguese and Croatian/Serbian.
-
📄 A Comparison of Transformer and Recurrent Neural Networks on Multilingual Neural Machine Translation
A comparative analysis of Transformer vs. recurrent architectures across bilingual, multilingual, and zero-shot translation — using professional post-edits and error categories, not just BLEU.
-
🏆 Best Paper Award at IWSLT 2017
Our paper “Improving Zero-Shot Translation of Low-Resource Languages” received the Best Paper Award at IWSLT 2017 in Tokyo.
-
📄 Improving Zero-Shot Translation of Low-Resource Languages
An iterative self-training procedure that improves zero-shot translation for low-resource languages — without any parallel data for the target direction.