CVPR 2019 about interpretable model update Feb/02/2919
-
Exploiting Kernel Sparsity and Entropy for Interpretable CNN Compression
-
Interpreting CNNs via Decision Trees
- related work: 1.gradient-based 2. up-convolutional nets feature map 3.semanctics of CNN -> object, parts, scenes, texture, materials, color. 1.diagnosis a pre-trained CNN
-
Interpretable and Fine-Grained Visual Explanations for Convolutional Neural Networks
Neural Architecture Search with Reinforcement Learning ICLR2017 Nov/20/2019
- Goal: Using RNN to design CNN and RNN architecture with Reinforcement Learning
- Reward is accuracy on test dataset.
- using Signmod to choose settings from sets.
- Heigh lights 1. variable length and structure. get some novel structures. 2. design structure automatically.
NLU = NLP + IR ?
update:Sep.2.2019
ERNIE 2.0: A CONTINUAL PRE-TRAINING FRAMEWORK FOR LANGUAGE UNDERSTANDING
1.word2vec 2.GloVe 3.Fasttext
- Byte Pair Encoding (BPE)
- WordPiece
- Unigram Language Model
1.The Corpus of Linguistic Acceptability
- Machine Translation
- single-sentence classification
- pairwise text classification
- pairwise text similarity
- relevance ranking
#ERNIE 2.0: A CONTINUAL PRE-TRAINING FRAMEWORK FOR LANGUAGE UNDERSTANDING