Applications of Deep Learning Alpha Go Google Translate Data Center Optimisation Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid November 23, 2016 Template by Philipp Arndt
Applications of Deep Learning Introduction November 23, 2016 FFR141 - Complex Systems Seminar Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid Applications of Deep Learning 2
Applications of Deep Learning Introduction AlphaGo Google s Neural Machine Translation (GNMT) Deep learning to control data centre cooling Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 3
Solving boardgames November 23, 2016 FFR141 - Complex Systems Seminar Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid Applications of Deep Learning 4
Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 5
Deep neural network Supervised learning Reinforcement learning from games of self-play Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 6
Deep neural network Supervised learning Reinforcement learning from games of self-play Monte Carlo simulation Policy network Value network Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 6
Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 7
Policy network: classifies promising positions Value Network: calculate estimates of winning Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 7
Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 8
99.8% Winratio against other go programs Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 8
Applications of Deep Learning Google s Neural Machine Translation (GNMT) Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 9
Google s Neural Machine Translation (GNMT) Introduction Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 10
Google s Neural Machine Translation (GNMT) Introduction There are flaws, BUT... September 27, 2016 GNMT announced Error reduction by 60% Bridging the Gap between Human and Machine Translation How? Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 11
Google s Neural Machine Translation (GNMT) Introduction Overview Models used so far GNMT Model Experiments and Results Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 12
Google s Neural Machine Translation (GNMT) Models used so far Overview Models used so far GNMT Model Experiments and Results Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 13
Google s Neural Machine Translation (GNMT) Models used so far Phrase-Based Machine Translation (PBMT) Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 14
Google s Neural Machine Translation (GNMT) Models used so far Phrase-Based Machine Translation (PBMT) Probability tables Linguistic properties Neural Machine Translation (NMT) Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 15
Google s Neural Machine Translation (GNMT) Models used so far Flaws of NMT Accuracy Speed / Computation Robustness Coverage Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 16
Google s Neural Machine Translation (GNMT) GNMT Model Overview Models used so far GNMT Model Experiments and Results Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 17
Google s Neural Machine Translation (GNMT) GNMT Model How does GNMT handle these problems? Speed / Computation Robustness Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 18
Google s Neural Machine Translation (GNMT) GNMT Model How does GNMT handle these problems? Speed / Computation Robustness Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 19
Google s Neural Machine Translation (GNMT) GNMT Model Architecture Parallelism Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 20
Google s Neural Machine Translation (GNMT) GNMT Model How does GNMT handle these problems? Speed / Computation Robustness Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 21
Google s Neural Machine Translation (GNMT) GNMT Model Segmentation: WordPiece Model (WPM) Abwasserbehandlungsanlage Abwasser behandlungs anlage sewage water treatment plant Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 22
Google s Neural Machine Translation (GNMT) Experiments and Results Overview Models used so far GNMT Model Experiments and Results Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 23
Google s Neural Machine Translation (GNMT) Experiments and Results Tests on Benchmark Sentence Pairs Workshop on Machine Translation (WMT) data set BiLingual Evaluation Understudy (BLEU) metric Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 24
Google s Neural Machine Translation (GNMT) Experiments and Results Tests on Benchmark Sentence Pairs Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 25
Google s Neural Machine Translation (GNMT) Experiments and Results Human Evaluation Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 26
Google s Neural Machine Translation (GNMT) Experiments and Results Bridging the Gap between Human and Machine Translation Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 27
Deep learning to optimize cooling of data centres Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 28
Deep learning to control data centre cooling Introduction Facts and definitions 1 Google search = keep a lightbulb going for 25s 40.000 searches/s PUE = Power Usage Efficiency Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 29
Deep learning to control data centre cooling Predicting PUE Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 30
Deep learning to control data centre cooling Predicting PUE 99.6% accuracy Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 31
Deep learning to control data centre cooling Predicting PUE Difficult to optimize efficiency Non-linear interactions between machines and environment Systems ability to adapt to operational changes Each facility has unique architechture Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 32
Deep learning to control data centre cooling Controling the data centre Results: 40% reduction in energy usage for cooling Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 33
Conclusions Beating human intuition in board games Solving Language translation tasks Outperforming human engineering abilities Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 34
Thank you for listening! Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 35
Discussion Questions What are the limitations of deep learning? Are there tasks for which the technique cannot be applied? Are there areas where deep learning should be used, but isn t? Who is responsible when a machine makes a critical error? For example: Who is responsible if an AI or machine causes a train to derail or fails to properly diagnose a patient? Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 36
Applications of Deep Learning References Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi. Google s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation. (2016) D. Bahdanau, K. H. Cho, Y. Bengio. Neral Machine Translation by Jointly Learning to Align and Translate. (2015) P. Koehn, F. J. Och, D. Marcu. Statistical Phrase-Based Translation. (2003) R. Sennrich, B. Haddow, A. Birch. Neural Machine Translation of Rare Words with Subword Units. (2016) S. Jean, K. Cho, R. Memisevic, Y. Bengio. On Using Very Large Target Vocabulary for Neural Machine Translation (2015) http://www.slideshare.net/nlab_utokyo/machine-translation-introduction http://slideplayer.com/slide/9202214/ https://research.googleblog.com/2016/09/a-neural-network-for-machine.html Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 37
Applications of Deep Learning References https://www.bloomberg.com/news/articles/2016-07-19/ google-cuts-its-giant-electricity-bill-with-deepmind-powered-ai http://www.theverge.com/2016/7/21/12246258/google-deepmind-ai-data-center-cooling https://googleblog.blogspot.se/2014/05/better-data-centers-through-machine.html https://deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-40/ https: //docs.google.com/a/google.com/viewer?url=www.google.com/about/datacenters/efficiency/ internal/assets/machine-learning-applicationsfor-datacenter-optimization-finalv2.pdf https://googleblog.blogspot.se/2009/01/powering-google-search.html http://www.internetlivestats.com/google-search-statistics/ Robin Sigurdson, Yvonne Krumbeck, Henrik Arnelid 38