Machine learning: the power and promise of computers that learn by example MACHINE LEARNING: THE POWER AND PROMISE OF COMPUTERS THAT LEARN BY EXAMPLE 1 Machine learning: the power and promise of computers that learn by example Issued: April 2017 DES4702 ISBN: 978-1-78252-259-1 The text of this work is licensed under the terms of the Creative Commons Attribution License which permits unrestricted use, provided the original author and source are credited. The license is available at: creativecommons.org/licenses/by/4.0 Images are not covered by this license. This report can be viewed online at royalsociety.org/machine-learning Cover image © shulz. 2 MACHINE LEARNING: THE POWER AND PROMISE OF COMPUTERS THAT LEARN BY EXAMPLE Contents Executive summary 5 Recommendations 8 Chapter one – Machine learning 15 1.1 Systems that learn from data 16 1.2 The Royal Society’s machine learning project 18 1.3 What is machine learning? 19 1.4 Machine learning in daily life 21 1.5 Machine learning, statistics, data science, robotics, and AI 24 1.6 Origins and evolution of machine learning 25 1.7 Canonical problems in machine learning 29 Chapter two – Emerging applications of machine learning 33 2.1 Potential near-term applications in the public and private sectors 34 2.2 Machine learning in research 41 2.3 Increasing the UK’s absorptive capacity for machine learning 45 Chapter three – Extracting value from data 47 3.1 Machine learning helps extract value from ‘big data’ 48 3.2 Creating a data environment to suport machine learning 49 3.3 Extending the lifecycle of open data requires open standards 5 3.4 Technical alternatives to open data: simulations and synthetic data 57 Chapter four – Creating value from machine learning 61 4.1 Human capital, and building skils at every level 62 4.2 Machine learning and the Industrial Strategy 74 MACHINE LEARNING: THE POWER AND PROMISE OF COMPUTERS THAT LEARN BY EXAMPLE 3 Chapter five – Machine learning in society 83 5.1 Machine learning and the public 84 5.2 Social isues asociated with machine learning aplications 90 5.3 The implications of machine learning for governance of data use 98 5.4 Machine learning and the future of work 100 Chapter six – A new wave of machine learning research 109 6.1 Machine learning in society: key scientific and technical chalenges 110 6.2 Interpretability and transparency 10 6.3 Verification and robustnes 12 6.4 Privacy and sensitive data 13 6.5 Dealing with real-world data: fairness and the ful analytics pipeline 114 6.6 Causality 15 6.7 Human-machine interaction 15 6.8 Security and control 16 6.9 Supporting a new wave of machine learning research 17 Annex / Glossary / Appendices 119 Canonical problems in machine learning 120 Glosary 12 Apendix 124 4 MACHINE LEARNING: THE POWER AND PROMISE OF COMPUTERS THAT LEARN BY EXAMPLE EXECUTIVE SUMMARY E evituce x yramus M enihca gninrael si a hcnarb fo laicifitra .tnacifingis I n ,erachtlaeh enihcam gninrael si ecnegiletni taht swola retupmoc smetsys gnitaerc smetsys taht nac pleh srotcod evig ot nrael yltcerid morf ,selpmaxe ,atad dna erom etaruca ro evitcefe sesongaid rof .ecneirepxe T hguorh gnilbane sretupmoc ot niatrec .snoitidnoc I n ,tropsnart ti si gnitropus mrofrep cificeps sksat , yltnegiletni enihcam eht tnempoleved fo suomonotua ,selcihev dna gninrael smetsys nac yrac tuo xelpmoc gnipleh ot ekam gnitsixe tropsnart skrowten sesecorp gninrael�yb morf ,atad rehtar erom .tneicife roF cilbup secivres ti sah eht gniwolof�naht .selur�demargorp-erp l a i t n e t o p o t te g r a t t r o p p u s e r o m y l e v i t c e f f e o t esoht ni ,den ro ot roliat secivres ot .sresu R tnece sraey evah nes gnitic xe secnavda A dn ni ,ecneics enihcam gninrael si gnipleh ni enihcam ,gninrael hcihw evah desiar sti ot ekam esnes fo eht tsav tnuoma fo atad seitilibapac sorca a etius fo .snoitacilpa elbaliava ot srehcraeser , yadot gnirefo wen I gnisaercn atad ytilibaliava sah dewola sthgisni otni , ygoloib ,scisyhp ,enicidem eht enihcam gninrael smetsys ot eb deniart no laicos ,secneics dna . erom a egral lop fo ,selpmaxe elihw gnisaercni retupmoc gnisecorp rewop sah detropus eht T eh U K sah a gnorts yrotsih fo pihsredael lacitylana seitilibapac fo eseht .smetsys nihtiW enihcam�ni .gninrael morF ylrae srekniht eht dleif flesti ereht evah osla neb cimhtirogla ni ,dleif�eht hguorht ot tnecer laicremoc ,secnavda hcihw evah nevig enihcam gninrael , s e s s e c c u s e h t U K sa h de t r o p p u s e c n e l l e c x e retaerg . rewop A s a tluser fo eseht ,secnavda ,hcraeser�ni hcihw sah detubirtnoc ot eht smetsys hcihw ylno a wef sraey oga demrofrep secnavda�tnecer ni enihcam gninrael taht ta ylbaeciton namuh-woleb slevel nac won esimorp hcus .laitnetop T eseh shtgnerts ni mrofreptuo snamuh ta emos cificeps .sksat hcraeser dna tnempoleved naem taht eht U K lew�si decalp ot ekat a gnidael elor ni M yna elpoep won tcaretni htiw smetsys desab eht erutuf tnempoleved fo enihcam .gninrael no enihcam gninrael yreve , yad rof elpmaxe E gnirusn eht tseb elbisop tnemnorivne rof ni egami noitingocer ,smetsys hcus sa esoht eht efas dipar�dna tnemyolped fo enihcam desu no laicos ;aidem eciov noitingocer gninrael liw eb laitnese rof gnicnahne ,smetsys desu yb lautriv lanosrep ;stnatsisa eht U s ’K cimonoce ,htworg ,gnieblew dna dna rednemocer ,smetsys hcus sa esoht , ytiruces dna rof gnikcolnu eht eulav fo gib‘ desu yb enilno .sreliater A s eht dleif spoleved . ’atad A noitc ni yek saera – gnipahs eht atad , rehtruf enihcam gninrael swohs esimorp , e p a c s d n a l g n i d l i u b , s l l i k s g n i t r o p p u s , s s e n i s u b fo gnitropus ylaitnetop evitamrofsnart dna gnicnavda hcraeser – nac pleh etaerc secnavda ni a egnar fo ,saera dna eht laicos .tnemnorivne�siht dna cimonoce seitinutropo hcihw wolof era MACHINE LEARNIN G : THE P O WER AN D PROMISE O F COMPUTERS THAT LEAR N BY EXAMPLE 5 EXECUTIVE SUMMARY The recent success of machine learning owes There is already high demand for people no small part to the explosion of data that is with advanced skills, with specialists in the available in some areas, such as image or field being highly sought after, and additional speech recognition. This data has provided resources to increase this talent pool are a vast number of examples, which machine critically needed. ‘No regrets’ steps in building learning systems can use to improve their digital literacy and informed users will also performance. In turn, machine learning help prepare the UK for possible changes in can help address the social and economic the employment landscape, as the fields of benefits expected from so-called ‘big data’, machine learning, artificial intelligence, and by extracting valuable information through robotics develop. advanced data analytics. Supporting the development of this function for machine There is a vast range of potential benefits learning requires an amenable data from further uptake of machine learning across environment, based on open standards industry sectors, and the economic effects and frameworks or behaviours to ensure of this technology could play a central role in data availability across sectors. helping to address the UK’s productivity gap. Businesses of all sizes across sectors need As machine learning systems become more to have access to appropriate support that ubiquitous, or significant in certain fields, three helps them to understand the value of data skills needs follow. Firstly, as daily interactions and machine learning to their operations. with machine learning become the norm for To meet the demand for machine learning most people, a basic understanding of the across industry sectors, the UK will need to use of data and these systems will become an support an active machine learning sector, important tool required by people of all ages which capitalises on the UK’s strength in this and backgrounds. Introducing key concepts area, and its relative international competitive in machine learning at school can help ensure advantages. The UK’s start-up environment this. Secondly, to ensure that a range of has nurtured a number of high-profile success sectors and professions have the absorptive stories in machine learning, and strategic capacity to use machine learning in ways that consideration should be given to how to are useful for them, new mechanisms are maximise the value of entrepreneurial needed to create a pool of informed users or activity in this space. practitioners. Thirdly, further support is needed to build advanced skills in machine learning. 6 MACHINE LEARNING: THE POWER AND PROMISE OF COMPUTERS THAT LEARN BY EXAMPLE EXECUTIVE SUMMARY T eh R layo S yteico detcudnoc hcraeser ot M enihca gninrael si a tnarbiv dleif fo dnatsrednu eht sweiv fo srebmem fo eht htiw�,hcraeser a egnar fo gnitic xe saera cilbup sdrawot enihcam .gninrael elihW rof rehtruf tnempoleved sorca tnerefid tsom elpoep erew ton erawa fo eht ,mret dna�sdohtem .snoitacilpa I n noitida ot yeht did wonk fo emos fo sti .snoitacilpa e s o h t s a e r a f o hc r a e s e r t a h t s s e r d d a y l e r u p T ereh saw ton a elgnis nomoc , weiv htiw lacinhcet ,snoitseuq ereht si a noitceloc ,sedutita evitisop�htob dna ,evitagen gniyrav fo cificeps hcraeser snoitseuq erehw gnidneped no eht secnatsmucric ni hcihw sergorp dluow yltcerid serda saera fo enihcam gninrael saw gnieb .desu gniognO cilbup dnuora�nrecnoc enihcam ,gninrael tnemegagne htiw eht cilbup liw eb tnatropmi ro stniartsnoc no sti rediw .esu S tropu sa eht dleif .spoleved rof hcraeser ni eseht saera nac erofereht pleh erusne deunitnoc cilbup ecnedifnoc M enihca gninrael snoitacilpa nac mrofrep lew ni eht tnemyolped fo enihcam gninrael ta cificeps .sksat I n ynam sesac ti nac eb de s u .smetsys T eseh saera edulcni cimhtirogla o t tn e m g u a n a m u h . s e l o r A hgu o h t l t i ra e l c � s i , s s e n t s u b o r,� , y c y a t v i i l r i p b a t ,e s r s p e r n e r t i n a i f t a h t s t n e m p o l e v e d n i en i h c a m g n i n r a e l l l i w ecnerefni fo , ytilasuac enihcam-namuh e g n a h c e h t dl r o w f o ,k r o w g n i t c i d e r p w o h siht ,noitcaretni dna . ytiruces liw dlofnu si ton ,drawrofthgiarts dna gnitsixe seiduts refid ylaitnatsbus ni rieht .snoitcejorp elihW gnirefo laitnetop rof wen sesenisub ro saera fo eht U K ymonoce ot ,evirht eht evitpursid laitnetop fo enihcam gninrael sgnirb htiw ti segnelahc rof , yteicos dna snoitseuq tuoba sti laicos .secneuqesnoc S emo fo eseht segnelahc etaler ot eht yaw ni hcihw wen sesu fo atad emarfer lanoitidart stpecnoc , fo rof ,elpmaxe ycavirp ro ,tnesnoc elihw srehto etaler ot woh elpoep tcaretni htiw enihcam gninrael .smetsys C lufera liw�pihsdrawets eb deden ot erusne taht ytivitcudorp�eht dnedivid morf enihcam gninrael stifeneb la . yteicos�ni MACHINE LEARNIN G : THE P O WER AN D PROMISE O F COMPUTERS THAT LEAR N BY EXAMPLE 7 RECOMMENDATIONS Recommendations EXTRACTING VALUE FROM DATA Creating a data environment In designing their studies, researchers should consider future potential uses of their data, to support machine learning and build in the broadest consents that are ethically acceptable, and acceptable to research Good progress in increasing the accessibility participants. of public sector data has positioned the UK as a leader in this area; continued efforts Research funders should ensure that data are needed in a new wave of ‘open data for handling, including the cost of preparing data machine learning’ by Government to enhance and metadata, and associated costs, such as the availability and usability of public sector staff, is supported as a key part of research data, while recognising the value of strategic funding, and that researchers are actively datasets. encouraged across subject areas to apply for funds to cover this. Research funders should In areas where there are datasets unsuitable for ensure that reviewers and panels assessing general release, further progress in supporting grants appreciate the value of such data access to public sector data could be driven management. by creating policy frameworks or agreements which make data available to specific users Extending the lifecycle of open under clear and binding legal constraints to safeguard its use, and set out acceptable data requires open standards uses. The UK Biobank demonstrates how such a framework can work. Government should New open standards are needed for data, further consider the form and function of such which reflect the needs of machine-driven new models of data sharing. analytical approaches. Continuing to ensure that data generated by The Government has a key role to play in the charity- and publicly-funded research is open creation of new open standards, for example by default and curated in a way that facilitates for metadata. Government should explore machine driven analysis will be critical in ways of catalysing the safe and rapid delivery supporting wider use of research data. Where of these to support machine learning in the UK. appropriate, journals should insist on this data being made available to other researchers in its original form, or via appropriate summary statistics where sensitive personal information is involved. 8 MACHINE LEARNING: THE POWER AND PROMISE OF COMPUTERS THAT LEARN BY EXAMPLE RECOMMENDATIONS Recommendations CREATING VALUE FROM MACHINE LEARNING Human capital, and building An analysis of the future data science needs of students, industry, and academia should skills at every level be undertaken to inform future curriculum developments. Schools need to ensure that key concepts in machine learning are taught to those who will To equip students with the skills to work with be users, developers, and citizens. machine learning systems across professional disciplines, universities will need to ensure Government, mathematics and computing that course provision reflects the skills which communities, businesses, and education will be needed by professionals in fields professionals should help ensure that relevant such as law, healthcare, and finance in the insights into machine learning are built into the future. Some exposure to machine learning current education curriculum and associated techniques will also be useful in many scientific enrichment activity in schools over the next activities. Professional bodies should work five years, and that teachers are supported in with universities to adjust course provision delivering these activities. accordingly, and to ensure accreditation In addition to the relevant areas of schemes take these future skills needs mathematics, computer science, and data into account. literacy, the ethical and social implications of In the short term, the most effective mechanism machine learning should be included within to support a strong pipeline of practitioners in teaching activities in related fields, such as machine learning is likely to be government Personal, Social, and Health Education. support for advanced courses – namely masters The next curriculum reform needs to consider degrees – which those working across a range the educational needs of young people of sectors could use to pick up machine learning through the lens of the implications of machine skills at a high level. Government should learning and associated technologies for the consider introducing a new funded programme future of work. of masters courses in machine learning, potentially in parallel with encouragement for approaches to training in machine learning via Massive Open Online Courses (MOOCs), with the aim of increasing the pool of informed users of machine learning. MACHINE LEARNING: THE POWER AND PROMISE OF COMPUTERS THAT LEARN BY EXAMPLE 9 RECOMMENDATIONS CREATING VALUE FROM MACHINE LEARNING (CONTINUED) Universities and funders should give urgent Machine learning and the attention to mechanisms which will help recruit Industrial Strategy and retain outstanding research leaders in machine learning in the academic sector. This As it considers its future approach to academic leadership is critical to inspiring immigration policy, the UK must ensure that and training the next generation of research research and innovation systems continue to leaders in machine learning. be able to access the skills they need. The UK’s approach to immigration should support In considering the allocation of additional the UK’s aim to be one of the best places PhD places, as announced in the Spring 2017 in the world to research and innovate, and budget, and new fellowships across subject machine learning is an area of opportunity in areas, machine learning should be considered support of this aim. a priority area for investment. Government’s proposal that robotics and AI Because of the substantial skills shortage could be an area for early attention by the in this area, near-term funding should be Industrial Strategy Challenge Fund is welcome. made available so that the capacity to train Machine learning should be considered a key UK PhD students in machine learning is able technology in this field, and one which holds to increase with the level of demand for significant promise for a range of industry candidates of a sufficiently high quality. This sectors. could be supported through allocation of the expected 1000 extra PhD places, or may UK Research and Innovation (UKRI) should require additional resources. ensure machine learning is noted as a key technology in the Robotics and AI Challenge area. In determining the shape and nature of DARPA-style challenge funding for research, Government should have regard to facilitating the spread and uptake of machine learning across sectors. 10 MACHINE LEARNING: THE POWER AND PROMISE OF COMPUTERS THAT LEARN BY EXAMPLE