Does AI need a body? - Dr. Oliver Bock

• Intelligence - nobody knows how to define it.
• Computation is not actually possible without a body. Even a computer needs some physical body.
• Human body:
  • 300m sensors of input,
  • 600 muscles of output.
• So we have this 300m-dimensional input, how do we interpret it? We need good priors, otherwise too many options. With no priors learning will be too slow.
• 3 insults
  • Heliocentrism,
  • Evolution,
  • Psychology,
• Now the 4th insult is coming: our intelligence is not special.
• We thought that intelligence is in the blood, then heart, then brain, but it's actually the body, and the whole society -- one human alone would not be that capable without civilization.
• AlphaGo might be very smart, but it can't move the stones by itself. But you all can, even though you can't play world class go.
  • Moravec's paradox -- what is easy for humans is hard for the machine.
• Perhaps if we want to replicate intelligence, we should start with things that are common to all intelligences, not just some humans. Let's start with interaction with the environment.
• Kitten experiment: can't train vision without locomotion.
• Conjecture: to really understand the natural intelligence, we need to start from the start - learn by interacting with the environment.
• Biological priors are based on coupling of body/mind/...
  • Experiment with priority of motion over color change
  • Biases: for example anchoring
• Intelligence = collection of these priors that is suitable for the situation.
  • We need to study the body and mind together.
  • Science of intelligence cluster at TU Berlin
• Cockatoo
  • brain is 6g
  • 100 minutes of exposure to the puzzle
  • they can also solve 4-digit locks
• Robots with machine learning suck compared to this, they are much less efficient... but why?
• Should we prioritise algorithms or data in our search for intelligence?
  • We need both: algorithms are like efficient priors but of course we need data to learn the environment.
  • Algorithms combined with deep learning are better than just DL. In particular, they are more data-efficient and generalize better. Basically this is not very surprising because the algorithm works as a very good prior.
    • Differentiable particle filter vs DL vs a combo in a maze.
• Grasping and manipulation complexity are correlated with brain size. Also when you look at FMRI scans, manual manipulation lights up the whole brain.
  • Grasping was very hard when roboticists were building stuff themselves but some simpler manuipulators combined with ML actually learn to pick up stuff in very effective, flexible and natural ways. Some part of the prior is the design of the manipulator -- we're using physical world to do part of our computation.
  • Funky sensors: microphone in the finger, then speaker + microphone

Questions

• Do we need a body if our environment is something like the internet? For example what if we make a bot that only needs to make money on the internet.
  • There are niches in the modern civilization where you can do things without a body, but they are rather narrow. If you want to interact with the physical world or maybe even understand it at all, you will need a body.
• Where do priors come from?
  • Evolution, society, etc... we don't know everything but we can study natural systems.

Links

• Science of Intelligence cluster at TU Berlin
• Oliver Brock's page on SoI website
• Email
• Differentiable particle filters paper