1 https://trallard.github.io/Talks/RSE-shefeld The state of machine - PowerPoint PPT Presentation

1

https://trallard.github.io/Talks/RSE-shef�eld The state of machine learning The state of machine learning RSE seminar, University of Shef�eld Tania Allard, PhD 2 . 1

Tania Allard Tania Allard Developer advocate Research Software Engineer Data expert  trallard  ixek 2 . 2

 ixek Machine learning Machine learning everywhere everywhere 3

 ixek Machine learning Machine learning everywhere everywhere So much that it is starting to not make sense anymore... like when you say a word 50 times in a row 3

 ixek For good or for bad it is everywhere: 4

 ixek For good or for bad it is everywhere:  Deployed in healthcare and warfare 4

 ixek For good or for bad it is everywhere: Deployed in healthcare and warfare  In the creative industry (from music to books)  4

 ixek For good or for bad it is everywhere: Deployed in healthcare and warfare  In the creative industry (from music to books)  Reading CVs and judging your creditworthiness  4

 ixek For good or for bad it is everywhere: Deployed in healthcare and warfare  In the creative industry (from music to books)  Reading CVs and judging your creditworthiness  Making us more Instagram worthy  4

 ixek The big players:  Apple  Facebook  Google IBM Intel  Microsoft Nvidia Open AI  Twitter 5

 ixek Machine learning generalised in two workflows Machine learning generalised in two workflows Model development (R&D) Model serving (production for customers consumption) 6

 ixek 7

 ixek What are these giants' issues? What are these giants' issues? 8

 ixek What are these giants' issues? What are these giants' issues? Mainly scale...in multiple areas 8

 ixek If we have a small team we have a smaller number of issues... right? 9

 ixek If we have a small team we have a smaller number of issues... right?  Small number of models to maintain 9

 ixek If we have a small team we have a smaller number of issues... right? Small number of models to maintain  People have the knowledge in their heads  9

 ixek If we have a small team we have a smaller number of issues... right? Small number of models to maintain  People have the knowledge in their heads  They have their own methods to track progress  9

 ixek That is the small team performance fallacy That is the small team performance fallacy We still need processes and best practices in place... so let me get back at this later 10

 ixek As the team As the team demand demand grows the problems grow grows the problems grow Increased complexity of data �ow  Larger number of work�ows  Managing complexity of �ows and scheduling becomes a nightmare  Resource allocation has to be on point  11

 ixek Serving models becomes harder Serving models becomes harder 12

 ixek

13

 ixek How do they serve How do they serve millions of millions of

customers across customers across the globe? the globe? 14

 ixek Three main players: Infrastructure / resources  Processes  People  15

 ixek

16

 ixek 17

 ixek Infrastructure as a code Infrastructure as a code 18

 ixek 19

 ixek Everything as a code Everything as a code Version control Less ambiguity on the con�gurations Shorter turnarounds Deterministic environments 20

 ixek Processes Processes 21

 ixek

22

 ixek Data and code as first class citizens Data and code as first class citizens

23

 ixek

24

 ixek People People Data scientist Data engineer ML Engineer 25

 ixek What does academia have to What does academia have to offer? offer?  Much more than you think 26

 ixek People People Researchers Research software engineers Librarians 27

 ixek Resources and Infrastructure Resources and Infrastructure We still need to �gure this out... it is pretty much an ad-hoc case 28

 ixek Processes Processes Scienti�c rigour Peer review Data management 29

 ixek Which areas could benefit from academic Which areas could benefit from academic collaborations? collaborations? 30

 ixek Meta-learning Meta-learning Humans learn across tasks (learn from experience)

31

 ixek If prior tasks are similar then we can carry prior knowledge 32

 ixek AlphaGo uses some sort of meta-learning 33

 ixek Algorithmic fairness Algorithmic fairness It has become increasingly important to ensure that models are making justi�ed calls that are free from unintended bias. 34

 ixek Algorithmic fairness Algorithmic fairness It has become increasingly important to ensure that models are making justi�ed calls that are free from unintended bias. The one way to make progress is through interdisciplinary collaboration 34

 ixek Towards model explainability Towards model explainability Address the trade-off between performance and interpretability 35

 ixek Reinforcement learning deadly triad Reinforcement learning deadly triad Following nature's paradigms RL agents receive awards and then learn to maximise success by performing optimal actions. 36

 ixek How to keep an algorithm learning if there are far too many potential variables or outcomes to be evaluated without being fed ridiculous amounts of data. 37

 ixek In brief In brief Focus on the 3 pillars: People  Infrastructure  Processes  38

Thank you Thank you  ixek  tania.allard@microsoft.com 39