Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

The future of AI: 12 possible breakthroughs in the next 5-10 years

12 ways in which the AI of 5-10 years time could be very different from today's. A presentation made by David Wood on 30th July 2020 to the Global Data Science and AI meetup.

A recording of this presentation can be accessed at https://www.youtube.com/watch?v=fILm4T6eXSE

For further discussion about the future of AI, and an opportunity to contribute to a new book on this subject, see the speaker's personal blog https://dw2blog.com/2020/07/31/the-future-of-ai-12-possible-breakthroughs-and-beyond/

  • Be the first to comment

  • Be the first to like this

The future of AI: 12 possible breakthroughs in the next 5-10 years

  1. 1. Artificial Intelligence in 5-10 years time: 12 ways it could be very different from today David Wood – @dw2 – London Futurists More than 200 public events since Feb 2008
  2. 2. @dw2 Page 2 More-of-the-same future Search for growth Incremental pipelineOperational excellence Game-changed future Creating a new future We all need to split our focus across three perspectives in parallel Disruptive change Present-day activities Fulfilling commitments
  3. 3. Vision: June 1998 Positive feedback cycle Spotted trends Anticipated convergence Patiently built a platform for collaboration
  4. 4. 1988 2013 Disruption: Slow & disappointing phase followed by fast & furious phase Disruption: Requires building not just new technology but also an ecosystem with feedback cycles Disruption: Happens in waves Each with new ideas, new tools, people Disruption: Outcomes are by no means inevitable! Human factors matter! Positive feedback cycle
  5. 5. AI in 2010 Classical AI – Expert systems Rules painstakingly hand-crafted by human experts Lots of ongoing progress: probabilistic reasoning… Aware of neural networks, but sceptical of them “Good old fashioned AI” (GOFAI) Perceptrons (1969) Marvin Minsky and Seymour Papert Deep Neural Networks Machine Learning (ML) Parameters automatically trained Unexpectedly swift progress in: Image recognition Language translation Speech recognition… Breakthroughs enabled by: Big Data (labelled) Improved hardware (GPUs, TPUs…) Improvements in algorithms Numerous innovations in design AI in 2015-2020
  6. 6. “Good old-fashioned neural networks” AI in 2020 Changes enabled by supply Supply of ideas Supply of human resources Supply of financial investment Changes driven by demand Huge profits to be made Numerous industries interested Algorithmic stock trading Healthcare + Medicine Gaming + Entertainment Engineering + Science… Control of the world at stake Deep Neural Networks Machine Learning (ML) Parameters automatically trained Unexpectedly swift progress in: Image recognition Language translation Speech recognition… Breakthroughs enabled by: Big Data (labelled) Improved hardware (GPUs, TPUs…) Improvements in algorithms Numerous innovations in design AI in 2025-2030
  7. 7. https://qz.com/1170185/the-master-algorithm-and-augmented- the-two-books-helping-chinas-xi-jinping-understand-ai/ China’s president Xi Jinping, bookshelves on New Year’s Day 2018 2017: China will become the leading AI power by 2030
  8. 8. People Technology Education Networks Tools Positive feedback cycle Entrepreneurs Engineers Scientists EducatorsDesigners Artificial Intelligence Deep Learning The acceleration of technology The acceleration of disruption
  9. 9. Positive feedback cycles Tools Machinery
  10. 10. Chemical reagents Synthetic chemicals Positive feedback cycles
  11. 11. Design, Manufacturing Computers Positive feedback cycles
  12. 12. Software tools (debuggers, compilers…) Software Positive feedback cycles
  13. 13. AI tools AI Positive feedback cycles
  14. 14. “Good old-fashioned neural networks” AI in 2020 Changes enabled by supply Supply of ideas Supply of human resources Supply of financial investment Changes driven by demand Huge profits to be made Numerous industries interested Algorithmic stock trading Healthcare + Medicine Gaming + Entertainment Engineering + Science… Control of the world at stake Deep Neural Networks Machine Learning (ML) Parameters automatically trained Unexpectedly swift progress in: Image recognition Language translation Speech recognition… Breakthroughs enabled by: Big Data (labelled) Improved hardware (GPUs, TPUs…) Improvements in algorithms Numerous innovations in design AI in 2025-2030
  15. 15. 1. Even bigger sets of data for ML to learn from Data generated and labelled by another AI Data cleaned by another AI before driving learning Qualitative change arising from quantitative change
  16. 16. 2. Transfer learning – enabling learning from small data Like brain trained by evolution to quickly learn new info Systems for unsupervised learning
  17. 17. 3. Systems that can self-learn natural language Start with small seed and it grows from there Iterative growth of “common sense” knowledge
  18. 18. 4. GANs: Generative Adversarial Networks Networks in adversarial relationship to each other Creativity generated via arms race
  19. 19. 5. Algorithms inspired by evolution Genetic algorithms…
  20. 20. 6. Systems inspired by new insights from neuroscience Like “neural networks” but with a more accurate correspondence 7. Neuromorphic computing New ideas in hardware as well as software
  21. 21. 8. Quantum computing Existing algorithms running faster Brand new algorithms possible
  22. 22. 9. Affective computing Artificial emotional intelligence 10. Sentient computing Consciousness designed in?! Real emotional intelligence?!
  23. 23. 11. AI that understands not just correlation but causation Beyond surface pattern recognition to deeper insights
  24. 24. 12. Intelligence emerging from decentralised network Simpler components combine to higher powers As targeted by SingularityNET
  25. 25. Potential timescale for major breakthroughs 1. Bigger sets of data 2. Transfer learning 3. Self-learned common-sense 4. Generative Adversarial Networks 5. Evolutionary algorithms 6. Neuroscience algorithms 7. Neuromorphic computing 8. Quantum computing 9. Affective computing 10. Sentient computing 11. AI that understands causation 12. Emergence from decentralised network My advice: keep an open mind Beware dogmatic sceptics Recall that “experts”, like other people, can become trapped into their current paradigm of thinking A single breakthrough from the list might open up many new insights Like quantum mechanics in 1925/26 Or like deep neural networks in 2012 Prepare NOW for potential breakthroughs And for potential bad consequences Consider “Anticipatory governance” Ensure sufficient research into the potential issues arising
  26. 26. Complex software with deeply complex (opaque) bugs Specification incomplete – unexpected behaviour arising Novel emergent capabilities and actions Software hacked or misled by adversarial attack Software unhackable – cannot be switched off Use by military Use by Big Brother surveillance Use by extreme capitalists over-prioritising profit, profit, profit Use by financial hackers or political hackers Use by suicidal fundamentalists
  27. 27. The Abolition of Aging: The forthcoming radical extension of healthy human longevity Transcending Politics: A technoprogressive roadmap to a comprehensively better future Sustainable Superabundance: A universal transhumanist invitation RAFT 2035: Roadmap to Abundance, Flourishing and Transcendence, by 2035
  28. 28. 1. Accountability: When a mishap arises from an AI solution, the company that provided the solution should not be able to shrug off any responsibility, e.g. by taking refuge in get-out clauses in one-sided license agreements 2. Data transparency: Users of AI systems should be made aware of potential limits and biases in the data sets used to train these systems 3. No hidden trade-offs: Trade-offs between different measures of fairness ought to be made explicitly, rather than implicitly (bear in mind that different ideas on fairness are often impossible to satisfy at the same time) 4. Explainability: Preference should be given to AIs whose outputs can be reviewed and explained, rather than users simply having to accept a prior track record of apparent success 5. Algorithm transparency: Users should be made aware of any latent weaknesses in the design characteristics of the model, including any potential for the system to reach unsound conclusions in particular circumstances 6. Anti-fragility: Users should be made aware of any potential disaster modes when solutions can tip over from being beneficial or occasionally mildly harmful, to having truly ruinous effects 7. Verifiability: Preference should be given to AI systems where it is possible to ascertain that the system will behave as specified, and where it is possible to ascertain whether the specification has significant holes in it 8. Security transparency: Users should be made aware of any risks of the AI being misled (e.g. by adversarial data) or having some of its safety measures being edited out or otherwise circumvented 9. Auditability: Preference should be given to AIs whose operation can be reliably monitored, by independent auditors, to ensure they are behaving according to prior agreements

×