> > > AGI's bottleneck must be in *learning*, anyone who focuses on something > else is barking under the wrong tree... >
Not just a bottleneck, it's the very definition of GI, the fitness / objective function of intelligence. Specifically, unsupervised / value-free learning, AKA pattern recognition. Supervision and reinforcement are simple add-ons. Anything else is a distraction. "Problem solving" is meaningless: there is no such thing as "problem" in general, except as defined above. > Now think about this: we already have the weapon (deep learning) which is > capable of learning *arbitrary* function mappings. > Yeah, we have some random hammer, so every problem looks like a nail. > We are facing a learning problem which we already know the formal > definition of. > No, you don't. There no constructive theory behind ANN, it's just a hack vaguely inspired by another hack: human brain. Which is a misshapen kludge, and this list makes it perfectly obvious. Sorry, can't help it. Anyway, this is my alternative: https://github.com/boris-kz/CogAlg > > On Mon, Jun 11, 2018 at 3:41 AM YKY via AGI <[email protected]> wrote: > On Mon, Jun 11, 2018 at 2:55 PM, MP via AGI <[email protected]> wrote: > >> Right. Most of them work off of a variant of depth first search which >> would usually lead to a combinatorial explosion, or some kind of heuristic >> to cut down on search time at some other cognitive expense... >> >> Not to mention most of them run off human made rules, rather than >> learning it for themselves through subjective experience. >> >> I highly doubt even Murray’s bizarre system subjectively learns. There >> are hand coded concepts in the beginning of his trippy source code. >> >> How can your system overcome this? How can it subjectively learn without >> human intervention? >> > > > AGI's bottleneck must be in *learning*, anyone who focuses on something > else is barking under the wrong tree... > > Now think about this: we already have the weapon (deep learning) which is > capable of learning *arbitrary* function mappings. We are facing a > learning problem which we already know the formal definition of. So we > just need to apply that weapon to the problem. How hard can that be? > > Well, it turns out it's very hard to understand the abstract (algebraic) > structure of logic, that took me a long time to master, but now I have a > pretty clear view of its structure. > > Inductive learning in logic is done via some kind of depth-first search in > the space of logic formulas, as you described. The neural network can also > perform a search in the weight space, maximizing some objective functions. > So the weight space must somehow *correspond* to the space of logic > formulas. > > In my proposal (that has just freshly failed), I encoded the formulas as > the output of the neural network. That is an example, albeit I neglected > the first-order logic aspects. > > Does this answer your question? > > And thanks for asking, because that helps me to clarify my thinking as > well... ☺ > > *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> Permalink > <https://agi.topicbox.com/groups/agi/T731509cdd81e3f5f-M91e23dbb79884c9e185a9424> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T731509cdd81e3f5f-M85c87b4f6aeac8dd7343b99c Delivery options: https://agi.topicbox.com/groups
