Tackling the Limits of Deep Learning
summary
Argues for one joint QA framework and shows concrete modules that work: a dynamic memory network with an episodic memory module that makes multiple attention passes to handle coreference and compositional questions; a pointer-sentinel mixture to copy rare or unseen words; tying input and output embeddings to cut perplexity; character-level embeddings to remove unknowns; a joint multi-task LSTM stack with successive regularization that achieves four state‑of‑the‑art results (chunking, parsing, relatedness, entailment); quasi‑RNNs that speed training by 1.4×–16.9×.
AI-written summary · about
chapters (26)
- 0:00Intro
- 0:32Happy Women's Day!
- 0:43Practical Deep Learning For Everybody
- 2:38The Limits of Single Task Learning
- 4:08Framework for Tackling NLP
- 4:21QA Examples
- 6:15First of Six Major Obstacles
- 8:45Tackling Obstacle 1: Dynamic Memory Network
- 13:13The Modules: Episodic Memory
- 13:41Related work
- 15:02Analysis of Attention for Sentiment
- 17:26Input Module for Images
- 18:26Attention Visualization
- 20:36Accuracy: Visual Question Answering
- 25:29Tackling Joint Training
- 28:57Training Details: Regularized Idea
- 30:06No Zero Shot Word Predictions
- 31:44Tackling Obstacle by Predicting Unseen Words
- 33:11Pointer Sentinel for Language Modeling
- 33:27Duplicate Word Representations
- 34:00Tackling Obstacle by Tying Word Vectors
- 34:20Language Modeling With Tying Word Vectors
- 35:20Stanford Question Answering Dataset
- 36:12Quasi-Recurrent Neural Network
- 36:59Q-RNNs for Language Modeling
- 37:39Q-RNNs for Sentiment Analysis
related talks
31mKeynote Q&A – Barbara Plank
01h 51mDeep Language Modeling for Question Answering using Keras
27mThe Cutting Edge of AI: From LLMs to Multimodal Intelligence
49mThe Building Blocks of Today’s and Tomorrow’s Language Models
42mReasonable AI
01h 31mNatural Language Understanding: Foundations and State-of-the-Art
26mTowards controllable language generation
23mUnveiling the Journey of NLP