WhoBuilt

Ex Machina's Scientific Advisor

Jun 28, 2017· 53 min· 4.3K views·Watch on YouTube

Speaker

Murray Shanahan, an AI researcher at DeepMind and Imperial College London, discusses AI, consciousness, and scientific advising on Ex Machina with a Y Combinator host.

The frame problem and changing fields

Shanahan recalls that AI was a niche field when he began his PhD in the 1980s. His thesis explored speeding up Prolog queries by reusing previously established relationships. He describes the frame problem as deciding what matters in a situation and avoiding needless reasoning about what stays the same, such as a room’s wall color when equipment moves.

Relevance in games and human-machine partnerships

Shanahan moved from classical, logic-based AI to computational neuroscience and back toward AI and machine learning. He sees the frame problem in systems that must identify useful information: in Space Invaders, for example, avoiding falling bombs matters more than the invaders’ colors or shapes. He also notes that Go players have used AlphaGo to explore strategies that conventional wisdom had rejected, and describes the potential for people and machines to complement each other.

Embodiment, consciousness, and Ex Machina

Shanahan says physical embodiment is central to human intelligence because brains support movement and interaction with a complex world. Alex Garland contacted him after reading his book Embodiment and the Inner Life; Shanahan read the script, discussed ideas with Garland, and made a few small corrections. He calls the film’s test a “Garland test”: Caleb sees Ava is a robot and must decide whether to treat her as conscious through their interactions.

Ambiguity, ethics, and the limits of AI forecasts

Shanahan describes a script direction for Ava’s escape scene that would have shown alien-looking visual effects from her point of view; the effect was cut, preserving ambiguity about her experience. He says Asimov’s laws are fictional devices and that current AI cannot comprehend such rules. He also cautions that reaching brain-scale computing power would not itself produce human-level intelligence, which may require unknown conceptual breakthroughs.

Deep reinforcement learning and other projects

Shanahan is interested in combining ideas from symbolic AI with deep reinforcement learning. He says DQN learned Atari games from pixels and trial and error, sometimes reaching human or superhuman performance, but learning was slow compared with how quickly people infer objects and rules. He also discusses advising a play about memory loss and working with the art collective Random International on Rain Room and a moving light sculpture that evokes a person.

Takeaways

  • Design learning systems to focus on information that affects the task, while setting aside irrelevant details.
  • Pair trial-and-error learning with representations or rules that help systems learn faster.
  • Treat brain-scale compute estimates as measures of hardware capacity, not evidence that human-level intelligence is near.
  • When assessing machine consciousness, distinguish observable behavior from certainty about subjective experience.
  • Use AI as a partner that can help people explore strategies they might not discover alone.

Mentioned in this talk

Search inside this talk

What was claimed 13

Show all 13 facts

Related talks

Ex Machina's Scientific Advisor, Murray Shanahan · WhoBuilt