A word is worth a thousand pictures: Convolutional methods for text
summary
Claims that convolutional nets beat RNNs on speed and deployability backed by concrete implementations in TensorFlow including DenseNet adaptation, deconvolution and dilated convolutions and a character‑level translation example. Reports parity with LSTM results while reducing model size from ~1,000,000 to 80,000 parameters and cutting training from two days to one hour, plus much cheaper real‑world inference; code and slides are provided via the uploader link.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:26Convolutions for text processing
- 3:33NLP tasks setup
- 7:53Challenges of natural language processing
- 8:15Convolution in NLP
- 10:44NLP system requirements
- 12:07Long-term text dependencies
- 14:21Recurrent neural networks explained
- 18:58Why convolutions beat RNNs
- 24:03Hierarchy in convolutions
- 27:15Residual connections expanding fields
- 31:26Production results
related talks
37mCode First, Math Later: Learning Neural Nets Through Implementation and Examples
56mTutorial on Deep Learning II
42mDeep Learning for Computer Vision
28mDemystifying Deep Neural Nets
01h 33mIntroduction to Convolutional Neural Networks using TensorFlow and Keras
55mA Tour of Deep Learning With C++
59mDeep Neural Networks for Video Applications
51mCognition and Computation