楽天技術研究所の次世代AI 技術への挑戦
3 likes1,273 views
The document discusses the transition from traditional deep learning and big data approaches towards next-generation AI technologies, emphasizing the importance of methods such as transfer learning and data augmentation. It identifies various applications and emerging paradigms, including reinforcement learning, multi-modal learning, and generative adversarial networks (GANs). The author highlights ongoing developments in AI capabilities related to human-like understanding and interaction.
More Related Content
![[DL輪読会]Fast and Slow Learning of Recurrent Independent Mechanisms](https://cdn.slidesharecdn.com/ss_thumbnails/20210604zhangxin-210604043141-thumbnail.jpg?width=560&fit=bounds)
Similar to 楽天技術研究所の次世代AI 技術への挑戦 (20)
楽天技術研究所の次世代AI 技術への挑戦
- 1. 次世代 への挑戦 と 時代の次へ Mar. 22nd , 2019 森 正弥 Masaya Mori 楽天株式会社 執行役員 楽天技術研究所 代表
- 2. 2 雑談 • ElMo のような大量データに基づく成果も確かに続いてい るが、Deep Learning と BigData だけの世界は終わり、次 世代AI への胎動が始まっている • DQN • Generative Adverasiral Network • Transfer Learning、Data Augmentation • BERT、BackTranslation
- 3. 3 Trend of Deep Learning and Bigdata is coming to an end. New paradigm is emerging.
- 4. 4 • ~ Dec. 2017: • Google • ~ Aug. 2018: • DeepL.com (they didn’t use Bigdata) • ~ Now: • Facebook AI Research
- 5. 5 スコア 出展:Google AI Blog https://ai.googleblog.com/2017/08/transformer-novel-neural- network.html (アクセス日 2019/03/26) Transformer の高スコア、しかし、Back Translation Model はこれをや すやすとこえた。
- 6. 6 Deep learning Small Dataset Big Dataset Great AI Deep learning So so AI Other Dataset 2nd Deep Learning Other so so AI Connect & Loop Learning Super AI Interactive Loop
- 7. 7 • GAN, Adversarial Network • AlphaGo Zero (Deep RL) • AICO (ad banner generator & CTR predictor) • Predictive Learning Real Example Generated Example Generator Noise Source Discriminator GAN Real Fake
- 8. 8 A Variety of Dataset Just bigdata
- 10. 10 Other projects will be organized under one vision. Pitari AIris RCP Creative AI Projects for the platform Fraud Detection Data-based Trading Delivery Optimization Language Learning Next Candidates
- 12. 12 Machine translation Voice recognition Automatic Speech Recognition for Product Voice Search Product Data science Sentiment Analysis + Category Grocery & food Subcategory Wine 我们真的很有诚意 了。 你说我一个老总都 亲自跑了好几趟了。 Machine translation
- 13. 13 Machine translation Automatic Speech Recognition for Product Voice Search Product Data science Sentiment Analysis + Category Grocery & food Subcategory Wine 我们真的很有诚意 了。 你说我一个老总都 亲自跑了好几趟了。 Machine translationAI organizing Chaotic Data AI understanding multi-languages AI understanding speechAI understanding Voice of Customers Voice recognition
- 14. 14 RIT Machine Translation matches Human Translation Rated by bi-lingual speakers on a 5-point scale for adequacy and fluency RIT Human Google for English to Spanish / Portuguese / French / Polish / German / Italian And, we‘re starting English to Japanese.
- 15. 15 Face recognition Contents Generation (Creative AI) Gender, Age, Emotion Recognition Product Recognition
- 16. 16 Face recognition Gender, Age, Emotion Recognition Product Recognition AI understanding Face AI understanding User visually AI understanding Object visuallyAI generating Digital Contents Contents Generation (Creative AI)
- 17. 17 SNEAKER SALE Up to 30% off Generation Prediction 30 Sneaker Sale Up to OFF % Sneaker up to 30% off Sale Sneaker up to 30% off Sale 30 Sneaker Sale Up to OFF % Image Segmentation Images Text Styles Assisting Graphic Design Process
- 18. 18 Mature-Level At leveraging Deep Learning Vision Voice (ASR) Language Voice (TTS) Big Gap, but bridgingSome Gap The strategies of each Program Management are different.
- 21. 21 Artificially increase the volume of the training dataset to improve accuracy. It is good for when data is insufficient, quality is low, or data is imbalanced to a specific category. Small Dataset Small Dataset Big Dataset Data Augmentation Data is enough Accuracy is increased Data is insufficient
- 22. 22 Data augmentation example : • Shifting vertical/horizontally • Invert vertically/horizontally • Enlarging/Minimizing • Rotating • Tilting vertically/horizontally • Cropping • Changing contrast Method Sample Image *Google Website (Machine Learning Crash Course)
- 23. 23 Data augmentation example : • Mixup is combination of two training data Method Sample Image *C. Summers et al., "Improved Mixed-Example Data Augmentation", 2018
- 24. 24 A model developed for a task is reused as the starting point for a model on a second task. By transferring, we can get improved result with small dataset. Concept Use Cases Big Dataset Small Dataset Pre-Training Re-Training Output Transfer Autonomous cars Realization of automatic cars by deep learning *Preferred Research (https://research.preferred.jp/2016/01/ces2016/) 43cm 20 cm (ex. Flower image) (ex. Animal image) TOYOTA / Preferred Networks
- 25. 25 With a pretrained model with Japan`s Ichiba data, transfer learning can help extract prospective customer, product recommendation and purchase prediction in US market . Japan EC data US EC data Pre-Training Transfer ・User purchase history ・Browsing history ・Review ・Advertisement click count ・Product search history… Prediction (in US market) ・Prospective customer extraction ・Product recommendation ・Purchase prediction … Re-Training
- 26. 26 Ensemble Learning Multi-modal Learning Reinforcement Learning Meta Learning
- 27. 27 Ensemble learning method is techniques that create multiple models and then combine them to improve prediction accuracy. Concept Use Cases Predict demand forecast with high accuracy by using multiple learning model. Manufacturer *FUJITSU website “FUJITSU Business Application Operational Data Management & Analytics” Prediction accuracyModel B Output Model C Model A Ensemble Normal Output Model FUJITSU Predict demand forecast
- 28. 28 Predict USD/JPY, NK225 and JGB on daily or weekly basis from past data by using machine learning. Ensemble learning is used as a method, and accuracy is improved. Index ・・・ Model B Model C Model A Past Data Future Prediction Accuracy Ensemble learning Nikkei 225 Bond Currency Index ・・・ Nikkei 225 Bond Currency Input Machine Learning Output Predict price
- 29. 29 Classify product catalog by using machine learning. Ensemble learning is used as a method, and accuracy is improved. Product catalog data Classification (Taxonomy) Input Output ・Title ・Product description etc. Model B Model C Model A Accuracy Ensemble learning (XGBoost) Machine Learning
- 30. 30 Detect merchants which can repay money from EC data with machine learning. Ensemble learning is used as a method, and accuracy is improved. EC data Credibility Score Input Output Tons of inputs • Can repay • Cannot repay Judge MerchantsModel B Model C Model A Accuracy Ensemble learning Machine Learning 30 Sneaker Sale Up to OFF %
- 31. 31 Multi-modal learning is a model to learn from multiple data source(text, image, voice, etc.). It is expected to high accuracy than model which learn from single source Concept Text Voice Image Multi-modal learning Increase accuracy of fraud item detection by using multimodal model : image, product name, description and price. EC Robotics Develop ASVR(Audio-Visual Speech Recognition), which has high noise-robust with combination of sound and video signals, Use Cases *Waseda University, Ogata tetsuya (https://pdf.gakkai-web.net/gakkai/ieice/icd/html/2017/view/I_01_02.pdf) *Mercari, Engineering Blog “https://tech.mercari.com/entry/2018/04/24/164919”, Text Multi- modal source Single source Voice Image (Video)＋ Honda Research Institute Mercari Image Text (Product name, Description etc. ) ＋
- 32. 32 Item Genre Classification : with Multi-modal learning Classifier based on CNN/RNN Final Result Text Data • Item Title • Item Description Image Data LSTM CNN
- 33. 33 Reinforcement Learning is machine learning on how software agents to take action in environment to maximize some notion of reward. Agents find optimal action model through trial and error. Software Agents User etc. Action Feedback Concept Use Cases Find optimal action model Alibaba has adopted reinforcement learning to improve commodity search EC Tech DeepMind’s AlphaGo beat champion in Go game. *Sigmoidal (https://sigmoidal.io/alphago-how-it-uses-reinforcement-learning-to-beat-go-masters/) *Analytics India Magazine (https://www.analyticsindiamag.com/how-alibaba-is-applying-virtual-taobao-to-simulate-e-commerce-environment/) Google Alibaba
- 34. 34 100% 50% 0% Time Trials 100% 50% 0% Time Trials A B C A wins! A B C Automatically A / B Test BANDIT 100% 50% 0% Time Trials A wins A B C Automatically＋Dynamic Dynamic BANDIT B wins A wins! Static Environment change Evolution!
- 35. 35 Example (Human case) Skill of riding bicycle = Stand up ＋ Ascend or descend a staircase etc. Meta learning is approach of learning to learn. It learn a variety of tasks from small amounts of data by utilizing past learning. • Learn task quickly from small amounts of data by utilizing past learning • Meta learning is deep learning algorithm close to human *Nikkei X TECH (https://tech.nikkeibp.co.jp/dm/atcl/mag/15/00189/00003/)