Human-AI Partnerships. Nick Jennings Vice-Provost (Research and Enterprise) & Professor of Artificial Intelligence

Transcription

1 Human-AI Partnerships Nick Jennings Vice-Provost (Research and Enterprise) & Professor of Artificial Intelligence

2 AI in the Movies 2

3 Stephen Hawking AI is Important The development of full AI could spell the end of the human race AI is our biggest existential threat Elon Musk Whoever becomes the leader in AI will become the ruler of the world Theresa May Vladimir Putin establish the UK as a world leader in AI 3

4 Why is AI so Popular (again)? Abundance of cheap processing power Increasing number of connected devices: on us and in world around us accessed at home, work and play These devices provide information: that measures ever more of everything that can be mashed-up in unforeseen ways All printed material in the world All words ever spoken by human beings Estimated internet traffic by end of petabytes (2 x bytes) 5 exabytes (5 x bytes) 1000 exabytes (1 x bytes) 4

5 Why is AI so Popular (again)? Humans alone cannot process volumes of data necessary to make informed decisions We need assistance, we need AI True across all information rich domains Machines alone can rarely do all of a complex activity Automation of sub-tasks Always activities that humans are best at? Build effective partnerships with technology Move away from humans being the masters and technology the slave Recognise complementary expertise and roles J. C. R. Licklider 5

6 Human-Agent Collectives 6

7 More Helpful Machines (Jennings et al., 2014) bring digital and physical worlds together make smart contributions to task at hand are better collaborators Intelligent Agents act autonomously and take the initiative Human-Agent Collectives combine human and machine intelligence 7

8 Flexible Autonomy neither agents, nor humans are always in charge Humans act with varying degrees of computer support Agents act with varying degrees of human support Relationship varies depending on context 8

9 Agile Teaming continually establish and manage collaborative relationships Groups of agents and humans: Come together when needed to achieve goals no individual can achieve in isolation Disband once cooperative action has been successful. 9

10 Incentive Engineering motivate by incentive, rather than diktat Design rewards so actions that are encouraged generate desirable outcomes. Consider what we need to know about humans to do this. 10

11 Accountable Information Infrastructure track information veracity and provenance Make sense of heterogeneous data that has varying degrees of reliability and accuracy. Allow veracity and accuracy of information to be confirmed and audited, while maintaining privacy and ethics standards 11

12 HACs in Disaster Response 12

13 Ethnographic Studies of First Responders Fort Widley, UK Command practices, information management and resource allocation Disaster City, Texas: Implicit coordination among team members Shared practices Diligence to safety Rescue Angel Thunder 2 week international multi-agency SAR exercise (planning to execution) Detailed planning procedures according to ISO 9001 Collaborative mapwork Information management Communications between Silver and Pathfinders Hampshire Fire Rescue Use of simulation software Fast pace & various stakeholders involved make information management and communication challenging 13

14 AI and Disaster Response AI summarizes huge volumes of info to highlight potential casualties Human responders add observations, AI updates its assessment to be consistent with this trusted info 14

15 Deployment in Nepal Rescue Global needed to determine placement of water filters within 50 mile radius of Kathmandu Crowd labelled 1200 Planet Labs satellite images Folded in OpenStreetMap building density data and inferred population density map using ORCHID data processing algorithms 15

16 AI and Disaster Response J1 16

17 Human-AI Partnership Using their experience, first responders determine which teams best suited to what areas AI computes draft schedule to minimise completion time, responders tweak as necessary AI monitors incoming information streams for things that may change existing priorities, Humans monitor response team progress for unexpected difficulties or delays 17

18 Impact on society 18

19 AI will Augment Jobs Many jobs have automatable sub-tasks partnership of complementary skills of humans and AI Software capable of analysing large volumes of legal data led to increase in number of legal clerks more cases taken Software capable of analysing large numbers of medical images will not lead to decrease in radiographers? more images examined 19

20 Job Automatability (Frey & Osborne, 2014) Not just manual jobs Amount of routine : 47% :49% 20

21 Job Automatability (OECD, 2016) 21

22 AI will Cause Job Losses Technological innovation often accompanied by job fears from Luddites onwards Jobs will disappear or require many fewer workers AI will impact many jobs blind to the colour of your collar Jerry Kaplan 22

23 AI will Create New Jobs (MITSloan Management Review, 2017) Jobs created 23

24 Key Challenges for AI Interpretability and transparency making workings and rationale for outputs understandable by humans ensuring systems are fair Verification and robustness methods for verifying systems that are learning based on their experiences knowing when machines don t know the answer Making machines better partners still often frustrating tools, unable to cooperate and unaware of their social context Accountability determining who is accountable when things go wrong 24

25 Conclusions Rises in processing power, memory and data availability let us solve problems we couldn t previously tackle AI essential for many complex problems wide range of impressive applications for narrowly defined tasks artificial general intelligence is a long way off AI will impact many aspects of our lives AI systems as partners leads to a rise in the humans, as well as the robots! 25

26 Human-AI Partnership: Cancer Diagnosis Researchers trained their models with millions of labelled images to find probability that a patch contains breast cancer 85% improvement for Human-AI Partnership (Wang et al., 2016) from PathAI 26

27 High Skill Jobs at Risk? 27