> We also know from 35 years of experience (beginning with Cyc) that logic based knowledge representation is not a path to AGI
I'm begging to differ! But then you already know that I do, I guess :-) On Mon, 30 Sep 2019 at 23:47, Matt Mahoney <[email protected]> wrote: > Boolean logic is a subset of neural networks. A single neuron can > implement any logic gate. Assume the output is clamped between 0 and 1. > > A and B = A + B - 1. > A or B = A + B. > Not A = -A + 1. > > But first order logic is not so simple. We also know from 35 years of > experience (beginning with Cyc) that logic based knowledge representation > is not a path to AGI, in spite of what seems to be a straightforward > approach not requiring a lot of computing power. > > I hope you understand why this is the case. You can't train logic models. > Language evolved to be learnable on slow, massively parallel networks. > Semantics comes before grammar. If you want to know what works, study text > compression. > > On Mon, Sep 30, 2019, 8:13 AM YKY (Yan King Yin, 甄景贤) < > [email protected]> wrote: > >> On 9/27/19, Steve Richfield <[email protected]> wrote: >> > YKY, >> > >> > The most basic function of neurons is process control. That is where >> > evolution started - and continues. We are clearly an adaptive control >> > system. Unfortunately, there has been little study of the underlying >> > optimal "logic" of adaptive control systems. >> > >> > I strongly believe that a different sort of "logic" is at work, and that >> > what we call "intelligence" is simply a larger adaptive control system >> > working according to that "logic". We are clearly more intelligent than >> > ants, but that is more quantatative than qualitative. >> > >> > Learning logic seems like a good idea, but you might want to reconsider >> the >> > logic you are learning. >> > >> > Steve >> >> >> Thanks for the comment. It is a very common objection indeed, and also >> has some good reasons behind it. >> >> From cognitive psychology most people tend to believe that the brain uses >> *model-based* reasoning instead of *rules-based* reasoning. We don't >> fully understand the brain's mechanism, but we may guess at some general >> principles. I think the brain uses some sort of neural representations, >> which are composed of neural "*features*", ie, certain patterns of >> neurons' activations. >> >> Each neuron is either ON or OFF or may be regarded to activate with a >> fuzzy truth value. Thus we can view each neural feature as a "micro" *logic >> proposition*. That creates a rough correspondence between neural >> representations and logic representations. >> >> Indeed, it is not so surprising, as we can express our thoughts into >> natural language with relative ease, and natural language has the structure >> of logic propositions. >> >> For example, the visual cortex can recognize images such as "cat" and >> "dog". Through a dynamical process, it recognizes the situation of "cat >> chases dog". This is likely represented by a juxtaposition of "cat", >> "chase", "dog" neural features. This is very similar to the >> *predicate-logic* expression: chase(cat, dog). >> >> So there may not be a huge gap between neural and logical >> representations. The next question is: How does the brain jump from one >> neural representation state to the next? In logic, this is achieved via >> *rules >> with variables and quantifiers*. >> >> At this point I am not so sure if the brain's mechanism is really similar >> to the logical mechanism. That's why I think you raised a good question, >> and I still don't have a good answer 😊 , but it is a good place to start >> thinking. >> >> So I assume the brain jumps from one neural representation to the next, >> and that such neural states are formed via juxtaposition of "*concepts*" >> (which are some neural activation patterns). One way that this may be >> different from logic rules is that the neural representations can be >> "distributive". >> >> I need to think about this more, but there may exist a rough >> correspondence between logic rules and neural state-transitions. >> >> >> *Artificial General Intelligence List <https://agi.topicbox.com/latest>* > / AGI / see discussions <https://agi.topicbox.com/groups/agi> + > participants <https://agi.topicbox.com/groups/agi/members> + delivery > options <https://agi.topicbox.com/groups/agi/subscription> Permalink > <https://agi.topicbox.com/groups/agi/T77af318d4abfa8a8-M76095de8167921d265cefbe0> > -- Stefan Reich BotCompany.de // Java-based operating systems ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T77af318d4abfa8a8-M723d6ae6316c4de68fb90834 Delivery options: https://agi.topicbox.com/groups/agi/subscription
