Deep Learning Tokyo 2016 参加メモ
Deep Learning Tokyo 2016に参加してきた。
Articulate and Actionable Deep Learning (Prof. Trevor Darrell)
- Articulate: Image Description
- Combine CNN visual model with RNN for sequential output
- Video Description
- sequence input & output
- Deep Compositional Captioner
- Learning to Compose Neural Networks for Question Answering
- Natural Language Object Retrieval
- Spatial Context Recuurent ConvNet (SCRC)
- 論文
- Deep Visuomotor Representations
Light Roasted Use of Caffe in Yahoo! Japan (Naoaki Yamashita)
- Case 1. Area Ratio Estimation (Face)
- 画像中のあるオブジェクト(ex.顔)の占める割合を求める
- 検索エンジンで顔の割合が多い画像のランクを上げるのに使用したり
- Case 2. Saliency Map
- Boolean Map Saliency
- Salient Object Subtiling (SOS)
- 画像中に顕著なオブジェクトが何個あるか数える
Machine Intelligence for Visual Recognition (Prof. Harada)
- Roi pooling
- Fast Distributed Machine Learning Framework using JavaScript
- Bag of Visual Words
- 空間情報が失われる
- featureをランダムに配置したのち、ジグソーパズルを解くように組み合わせて元の画像情報を再構成
- image reconstruction from Bag of Visual Words
- 2 dimensional LSTM
Deep Learning for News Recommendation (Shumpei Okura)
- 記事のベクトル化
- De-noising Auto-Encoder
- (0, 1)のBoW -> 一部隠す -> 低次元にencode -> 高次元にdecode
- 重複排除にも使う
- De-noising Auto-Encoder
- ユーザ特徴のベクトル化
- 閲覧履歴をRNNにかける
- Training with click feedback by BPTT
Deep Learning in real world; Automobile, Robotics, Bio Science (Daisuke Okanohara)
- Industrial, Device方面にPositioning
- Anomaly Detection from Sensors
- Deep Learningを使った異常検知
- Trained machine can teach other machine
- Distillation (Hinton+ 2015)
- Privileged Information (Vapnik+ 2014, 2015)
Overview of Chainer and Its Features (Seiya Tokui)
- Chainer is a deep learning framework for researcherd with high flexibility and easiness to write NNs
- Define and Run vs Define by Run
- Define and Run (most framework)
- 最適化しやすい
- Portability高い
- Define and Run (most framework)
-
- Define by Run (Chainer, autograd)
- 構造をdynamicに変更できる
- Define by Run (Chainer, autograd)
- "device-agnostic codes"