7
$\begingroup$

I am a strong believer of Marvin Minsky's idea about Artificial General Intelligence (AGI) and one of his thoughts was that probabilistic models are dead ends in the field of AGI.

I would really like to know the thoughts and ideas of people who believe otherwise.

$\endgroup$
  • 8
    $\begingroup$ Where did Minsky say that probabilistic models are dead ends? Can you provide a reference, interested. $\endgroup$ – Gottfried William Dec 14 '16 at 18:09
2
$\begingroup$

When considering effective approaches to AGI, one must extrapolate outwards to the types of modelling (and therefore inputs) that would be necessary to achieve any general utility. One consideration might be the fundamental "building blocks" of our physical world, and understanding the movements of these, can lead to accurate predictions of (all) occurrences. These fundamental elements are called (generally) subatomic particles and are anything but discrete values. In quantum field theory, the more accurate you are able to measure position, the less accurate you may know a quarks momentum (and vice versa). Our world, at the most fundamental layer, is probabilistic when observed. This is all not to say that an understanding of quantum mechanical kinematic descriptions is the only methodology to achieve true AGI, but to say probabilistic models, and therefore uncertainty, is a dead-end seems radically inaccurate.

That said, Dr.Minsky didn't really feel probabilistic models were dead ends. The emerging view in the field, one that Dr.Minsky urged for years, is that connectionism alone couldn't exclusively lead to AGI due to its uniformed structure. If you are unaware, connectionism is the concept of creating models around discrete units of representation (neurones in a neural net for example). You see the problem we have identified isn't that probabilistic models are inaccurate, it's that our current approach doesn't express the biological realism necessary for AGI (although sufficient for specific intelligence).

[I briefly worked with Dr.Minsky at the AI Lab before his passing last year, a hilarious man and brilliant scientist.]

|improve this answer|||||
$\endgroup$
  • 1
    $\begingroup$ I don't think "quantum mechanical kinematic descriptions" have anything whatsoever to do with achieving true AGI, but I'd love to see you expand on the second paragraph. $\endgroup$ – BlindKungFuMaster May 30 '17 at 8:05
2
$\begingroup$

I think Minsky deprecated the suggestion that probabilistic models could be surrogates for component models for intelligence that he suggested were grounded in principles and processes that interact (i.e. Society of Mind). But I don't believe he ever referred to probabilistic models as dead ends. All intelligence models must employ awareness of likelihoods and exceptions. Many probabilistic techniques like bayesian inference and markov chains likely will prove essential to the eventual mechanisms of any AGI.

However I would agree with you that Minsky would be unhappy at the prospect of designing AGI systems using techniques like neural nets that 1) are shaped only by supervised training examples, and 2) are unable to explain their reasoning or revise it directly and efficiently. These approaches (disembodied probabilities) do not model thinking but only the outcome of thinking. That's akin to not understanding the question but guessing correctly anyway, which is certainly useful, but it's also pretty uninteresting for someone like Minsky who wanted to understand the thought process, not just build a Magic 8 Ball.

How would Minsky respond to recent advances in deep nets like variational autoencoders or adversarial nets or compound hierarchies of nets or end-to-end nets that implement broad skills? Assuming they can scale up to implement AGI successfully, I believe he'd be rather unhappy with them. The thought that his Society of Mind could turn out to me no more than a society of ten(?) kinds of deep nets that match patterns probabilistically and interact combinatorically -- I think this would be deeply disappointing to someone who spent a lifetime imagining and refining axioms and theories and proofs of cognition, only to find that the exquisite minds of Einstein and Beethoven might be little more than largely reactive engines highly attuned to variations in rote.

|improve this answer|||||
$\endgroup$
  • $\begingroup$ Hey, @Randy. I think the only plausible interpretation of it is that AGI design process isn't just a mathematical optimization problem, which is what most if not all ML models do, they optimize the set of actions and preferences. To some extent, the way we play with numbers without truly understanding them what they mean is, in fact, a dead end. Regardless, thank you for your answer. Best! $\endgroup$ – Parth Raghav Jul 16 '17 at 10:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.