MaCro Philosophy

Kinds of Intelligence Symposium at NIPS

This year I helped organise the Kinds of Intelligence Symposium at NIPS. Videos of (most of) the talks are now available. I've collected some common themes of the symposium. When you mouse over the themes, the relevant speakers are highlighted based solely on the topics covered in their talks (not their overall opinions). Video links to the talks and panel discussions, with a few brief sentences from my notes, are below.

Lucia Jacobs
Lucia Jacobs
Alison Gopnik
Alison Gopnik
Demis Hassabis
Demis Hassabis
Josh Tenenbaum
Josh Tenenbaum
Gary Marcus
Gary Marcus
Katja Hofmann
Katja Hoffman
We need to look beyond fully-formed humans to understand intelligence.
  • At the evolutionary history of plants and animals.
  • At cognitive development through childhood.
The amount of prior knowledge we put into our systems is important.
  • New advances require less built-in knowledge (and less search).
  • Innate knowledge is fundamental.
There are important differences between human intelligence and Artificial General Intelligence (AGI).
  • Games provide a way to compare human and artificial intelligence.
  • Language will be an important component in building intelligence.
Lucia Jacobs
Lucia Jacobs - The evolution of navigation: "Everything is space and odour"
  • To understand intelligence we must go back to first principles (of how it evolved). On this planet that is olfaction.
  • Intelligence in plants. The Mimosa plant habituates to being dropped by closing its leaves in anticipation of landing.
  • Intelligence from navigation. 51 million years ago the Cambrian explosion/bloodbath. Evolving complex brains via need to map environment as prey/predator (using olfaction).
Closing Remark: "The act of smelling something, anything, is remarkably like the act of thinking itself." (Lewis Thomas)
Alison Gopnik
Alison Gopnik - The distinctive intelligence of young children.
  • If you want to make intelligent AI, model it on human development.
  • Strong correlation between intelligence and time spent in childhood as a species. Humans have an incredibly unusual and costly extended childhood.
  • Childhood is evolution's way of resolving explore/exploit trade-offs and performing simulated annealing.
Closing Remark: If you want an AI that can learn as well and intelligently as humans, it should look more like [a child] than even those of us that are in this audience.
Demis Hassabis
Demis Hassabis - Learning from first principles: Introducing AlphaZero.
  • 4 dimensions of intelligence: Learning vs handcrafted, General vs Specific, Grounded (in reality) vs Logic-Based, Active vs Passive.
  • Amount of search per decision progressing towards humans (GM considers 10's, AlphaZero tens of thousands, Stockfish tens of millions).
  • "Deep Blue was the end; AlphaGo is a beginning" (Garry Kasparov).
Closing Remark: Human ingenuity augmented by AI will unlock our true potential.

Panel Discussion 1

Josh Tenenbaum
Josh Tenenbaum - Types of intelligence: why human-like AI is important.
  • We have AI technologies but (from a cognitive systems point of view) no real artificial 'intelligence'.
  • Intelligence is not just about pattern recognition, it is also about modeling the world. It is not just about calculating one kind of function really, really well, it is thinking about the problem we want to solve.
  • Build (not learn) commonsense core. Use commonsense core to learn language. Use language to learn everything else.
Closing Remark: Cognitive development provides a roadmap, and the whole emerging AI landscape provides a powerful set of tools which, as we start to bring them together, are beginning, but just beginning, to let us solve this.
Gary Marcus
Gary Marcus - The road to artificial general intelligence.
  • Thoughts take place within a mental language - but current direction of AI research is away from this.
  • Cognition = f(a,k,e) - a = innate algorithms, k = innate knowledge, e = experience. AlphaZero low k, but high a.
  • There is a lot of innateness in every neural network. Architecture choices, Network dynamics, Representational choices, Data choices, etc etc.
Closing Remark: The challenge of reusability ... is the real challenge that the field should be focussing on.
Katja Hofmann
Katja Hofmann - Video games and the road to collaborative AI.
  • Games a good tool to solve the problem of comparing AI and humans on an equal setting.
  • But skill at artificial games very different from human-like intelligence. Can easily demonstrate super-human performance without heading towards human-like AGI.
  • Ability on a broad range of tasks does not necessarily mean anything like human-like intelligence.
Closing Remark: Video games are a rich environment and flexible environment for measuring intelligence, but are not necessarily aligned well with progress towards human-like intelligence.

Panel Discussion 2

More information about the event can be found at its webpage.

Blog Homepage