Machine learning

1. Machine Learning - 1 - Prabhat HPC User Forum April 12, 2016

2. Slide Courtesy of Nervana Systems

3. (2012) This is all great, but… •  Is Machine Learning relevant to science? •  Why should HPC facili;es care about Machine Learning, Deep Learning, Sta;s;cs? - 4 -

4. (2012) This is all great, but… •  Is Machine Learning relevant to science? –  Success stories are for images and audio, but how about scien=ﬁc data? •  Why should HPC facili;es care about Machine Learning, Deep Learning, Sta;s;cs? –  Our applied mathema=cians are content with formula=ng and solving PDEs –  The NNSA folks care about Uncertainty Quan=ﬁca=on –  Our data ‘analy=cs’ folks are happy dealing with meshes, computa=onal geometry, topology - 5 -

5. (2016) The writing is on the wall •  O(B) $ worth of investment by industry •  Machine Learning and Sta;s;cs are established as key disciplines for this decade –  Deep Learning has taken oﬀ as the most promising ML technique - 12 -

6. (2012) Revisited.. •  Is Machine Learning relevant to science? •  Why should HPC facili;es care about Machine Learning, Deep Learning, Sta;s;cs? - 13 -

7. Astronomy Physics Light Sources Genomics Climate (2010-2016): The Rise of Data-Intensive Science

8. - 15 -

9. - 17 -

10. - 18 -

11. - 19 -

12. - 20 -

13. - 21 -

14. - 22 -

15. - 23 -

16. 4 V’s of Scientiﬁc Big Data - 24 - Science Domain Variety Volume Velocity Veracity Astronomy Mul=ple Telescopes, mul=-band/spectra O(100) TB 100 GB/night – 10 TB/night Noisy, acquisi=on artefacts Light Sources Mul=ple imaging modali=es O(100) GB 1 Gb/s-1 Tb/s Noisy, sample prepara=on/acquisi=on artefacts Genomics Sequencers, Mass- spec, proteomics O(1-10) TB TB/week Missing data, errors High Energy Physics Mul=ple detectors O(100) TB – O(10) PB 1-10 PB/s reduced to GB/s Noisy, artefacts, spa=o- temporal Climate Simula=ons Mul=-variate, spa=o-temporal O(10) TB 100 GB/s ‘Clean’, need to account for mul=ple sources of uncertainty

17. Does Machine Learning matter? •  Is Machine Learning relevant to science? –  Yes! •  Why should HPC facili;es care about Machine Learning, Deep Learning, Sta;s;cs? –  Analy=cs is the key step for gaining scien=ﬁc insights –  The nature of ques=ons in data-intensive science are inferen=al –  Sta=s=cs and Machine Learning deal with inference in presence of noise and errors - 25 -

18. Creating a catalog of all objects " in the Universe 1 Top 10 Data Analytics Problems

19. Fundamental Constants of Cosmology 2 Top 10 Data Analytics Problems

20. Characterizing Extreme Weather in a Changing Climate 3 Top 10 Data Analytics Problems

21. Knowledge Extraction from " Scientiﬁc Literature 4 Top 10 Data Analytics Problems

22. Understanding Speech Production 5 Top 10 Data Analytics Problems

23. Quantitative and Predictive Biology 6 Top 10 Data Analytics Problems

24. Understanding the Genetic Code 7 Top 10 Data Analytics Problems

25. 8 Top 10 Data Analytics Problems Personalized Toxicology

26. Designer Materials 9 Top 10 Data Analytics Problems

27. Fundamental Constituents of Matter 10 Top 10 Data Analytics Problems

28. Towards Synthesis (and maybe Convergence) •  What is the landscape of Machine Learning problems in science? –  Bewildering array of taxonomy and domain-speciﬁc terminology •  What are the key computa;onal mo;fs? –  Need to have a produc=ve conversa=on with HPC soaware, hardware vendors - 36 -

29. ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ Classiﬁca;on Regression Clustering Dimensionality Reduc;on Inference Model Es;ma;on Design of Experiments Seman;c Analysis Feature Learning Anomaly Detec;on Astronomy Cosmology Climate Systems Biology Neuroscience EM/X-Ray Imaging Mass-spec Imaging Personalized Toxicology Materials Par;cle Physics

30. ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ Classiﬁca;on Regression Clustering Dimensionality Reduc;on Inference Model Es;ma;on Design of Experiments Seman;c Analysis Feature Learning Anomaly Detec;on Astronomy Cosmology Climate Systems Biology Neuroscience EM/X-Ray Imaging Mass-spec Imaging Personalized Toxicology Materials Par;cle Physics

31. Hardware Op=mized Libraries Big Data Mo=f Scalable Algorithms Scien=ﬁc Analysis Science Applica=ons Astronomy, Cosmology, Climate, BRAIN, BioImaging, HEP Pafern/ Anomaly Discovery Large Scale Inference Clustering, Dimensionality Reduc=on Data Fusion Genome Assembly Dense/Sparse Linear Algebra Op=miza=on (Stochas=c) Randomized Linear Algebra Many-Core Chipset, Deep Memory Hierarchy, Reducing Data Movement, Power Eﬃciency Graph Methods (BFS, DFS,…) MapReduce ScaLAPACK, BLAS PCL-DNN TensorFlow SpearMint RandLA TECA GraphLab Deep Learning Stochas=c Varia=onal Inference Distributed MCMC Direct Graph Kernel Computa=on Sparse Coding DBSCAN CUR/CX Machine Learning Research Strategy

32. Hardware Op=mized Libraries Big Data Mo=f Scalable Algorithms Scien=ﬁc Analysis Science Applica=ons Astronomy, Cosmology, Climate, BRAIN, BioImaging, HEP Pafern/ Anomaly Discovery Large Scale Inference Clustering, Dimensionality Reduc=on Data Fusion Genome Assembly Dense/Sparse Linear Algebra Op=miza=on (Stochas=c) Randomized Linear Algebra Many-Core Chipset, Deep Memory Hierarchy, Reducing Data Movement, Power Eﬃciency Graph Methods (BFS, DFS,…) MapReduce ScaLAPACK, BLAS PCL-DNN TensorFlow SpearMint RandLA TECA GraphLab Deep Learning Stochas=c Varia=onal Inference Distributed MCMC Direct Graph Kernel Computa=on Sparse Coding DBSCAN CUR/CX Machine Learning Research Strategy

33. Machine Learning: Challenges •  Cultural –  ML doesn’t cleanly ‘ﬁt’ within Computer Science or Applied Math –  Sta=s=cs, CS (Machine Learning, HPC) taxonomy –  Mindshare •  Afrac=ng the best academic and industry talent is hard •  Technical –  Big Data ecosystem has evolved independently of HPC –  Aspira=ons of Convergence (Soaware, Hardware) •  HPC ins=tu=ons need to do a befer job of characterizing their Data Analy=cs requirements - 41 -

34. Machine Learning: Opportunities •  HPC community is uniquely posi;oned –  Storage and Compute Hardware –  Meaningful scien=ﬁc problems •  SoWware (Research and Produc;on) is wide open •  Most exci;ng discoveries happen at the intersec;on of domain sciences and methods –  We don’t know the limits of Deep Learning methods - 42 -

35. - 43 -

36. Thanks! - 44 - We are hiring: Big Data Architects Big Data Engineers Data Scien;sts Post-docs, interns Contact: prabhat@lbl.gov

Machine learning

You are reading a preview.

Check these out next

Machine learning

More Related Content

Viewers also liked (20)

Similar to Machine learning (20)

More from inside-BigData.com (20)

Recently uploaded (20)