Quoc Le, Software Engineer, Google
summary
Proposes a practical sequence-to-sequence encoder–decoder using recurrent neural networks that encodes a variable-length sentence into a 1000-dimensional vector and decodes it word-by-word with beam search. Trains on the WMT benchmark and reaches performance comparable to a decade-plus engineered MT system after about a year of work, can regenerate 20–30 word inputs, projects sentence vectors with PCA to show semantic clustering (subject/object, synonyms), and extends to QA and image captioning with promising results.
AI-written summary · about
chapters (4)
related talks
26mQuoc Le, Software Engineer, Google
26mUsing deep learning in natural language processing
39mDeep Learning for Machine Translation: a paradigm shift
36mNeural Networks
53mTensorflow, deep learning and modern RNN architectures, without a PhD
39mTackling the Limits of Deep Learning
26mTowards controllable language generation
42mAttention and Memory in Deep Learning Networks