Data scientist. AI Researcher

There is no magic

Precedents

  • Failure cases:
    • Shamans claimed to cure diseases.
    • Witches casted love spells.
    • Dowsers attempted to find water
    • Countless other people claimed to be able to perform a certain task and failed.
  • Success cases:
    • Modern astronomers predict solar eclipse.
    • Engineers design cars.
    • Many others perform a task reliably.
  • The difference:
    • The former group can’t explain why they should succeed or doesn’t even attempt to.
    • The later group can prove that their success must follow from an established body of knowledge.
  • Proofs can take many different forms but they are all independent of experiments.
    • To borrow the language of logics, they are “syntactic” instead of “semantic”.
    • In other words, they fall into the realm of “reason” as opposed to “experience”.
    • Showing that you succeed 99% (or even 100%) of the times doesn’t amount to a proof. Such activities are experiments and the results are sensitive to the settings and assumptions of that particular experiment.

The rule

  • There is no magic in this world.
  • If an ability can’t be proved (i.e. doesn’t follow from its description), it is likely not a real ability.
  • The performer might succeed by luck but they will fail when more cases are tested and they will fail in a systematic way.
  • The only way to perform a task reliably is to construct a procedure that can provably perform it.

The claims

  • Modern neural networks are the equivalent of shamans, witches, and dowsers: they succeed for the wrong reasons and fail systematically.
  • Only (future) neural networks that we can prove to have a ability truly have that ability.
    • Distributed representations cannot be proven and therefore will never be reliable.

Supporting evidence

  • Current neural networks falls prey to adversarial examples and fooling examples.
  • Current neural networks are strongly activated to random samples.
  • Current neural networks don’t work for out of distribution examples.

The solution

  • The only way to prove the ability of a neural network is to identify and prove the ability of its individual (or small group) of neurons.
  • We need to create “grand-mother neurons” or small groups of neurons that can be proven to perform a functionality.
  • We need to apply logics and math to proving the functionality of artificial neurons (repeat: calculating evaluation metrics is not proving).