Deep Language Modeling for Question Answering using Keras
summary
Demonstrates how to modify Keras to reproduce the published QA models (Feng et. al. 2015 and Tan et. 2016), showing exact layer choices, loss and optimizer tweaks, and deployment notes. Shows word embeddings, vanilla RNN vs LSTM training where LSTM fixes vanishing gradients and drives error far lower, and an attention-wrapped LSTM that highlights the answer token. Provides runnable notebooks and code, trains on GPUs and cloud like Google Compute Engine, and reports ~60% top‑1 accuracy on the published answer‑selection benchmarks.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 4:54Deep learning for data modeling
- 15:43Moving from Theano to Keras
- 21:52Word embeddings in NLP
- 39:53RNNs for sequence modeling
- 1:04:22Vanilla RNN in Keras
- 1:17:33Vanishing gradients and LSTMs
- 1:26:09RNNs and LSTMs with Keras
- 1:33:22Attention mechanisms in Keras
- 1:43:32Q&A
related talks
39mTackling the Limits of Deep Learning
39mDeep Learning vs. Conventional Machine Learning
01h 30mIntroduction to Deep Learning & Natural Language Processing
33mBuilding the next-generation Conversational AI with Python and Deep Learning
34mIs That a Duplicate Quora Question?
42mDevOps & Microservices: “Delivering Machine Learning models early and often.”
31mMachine Learning-Enabled Chatbots
45mBuild a Q&A Bot with DeepLearning4J