This version is an improvement -- basically I get where you are headed: "The gist of our theory is that Deep Learning provides us with neural networks ... that serve as the proof mechanism of logic..."
and why not? On 7/19/21, immortal.discover...@gmail.com <immortal.discover...@gmail.com> wrote: > On Sunday, July 18, 2021, at 8:59 PM, YKY (Yan King Yin, 甄景贤) wrote: >> Final version of my paper (corrected a lot of inaccuracies): >> https://drive.google.com/file/d/1P0D9814ivR0MScowcmWh9ISpBETlUnq-/view?usp=sharing >> You seem to mix together > 1) BERT's prediction of next word in text > with > 2) prediction of next item in IQ tests > but these two are not exactly the same...? > You may argue that humans can do both, > and indeed an AGI should be able to do that too. > Typically, it would require multiple steps of reasoning, > you were just looking at 1 layer of Transformer or Attention Mechanism, > that corresponds to a single step of inference. > Because prediction is prediction, the use of patterns is all that AGI does. > There is different pattern finders for different types of problems. Saying > to a robot "if you (see 1 cat) or (see 1 dog and a plane) walk forward else > walk backwards" is a sort of match and then prediction as output. > Nonetheless your paper is too un-unified, I'd like to see something way more > simpler and evil and easy that explains everything about AGI. You can do a > lot of damage with just 100 words and commonly used words instead of "topos > theory". If we want to unite ideas to become openAI2, explain your work > using common words/ sentences. > > Question: > I've asked a few people a question around forums and either they are too > busy Transforming, or can't answer the question properly because really they > didn't invent the Transformer architecture and are just users, lots of users > albeit. I saw DALL-E/ IGPT predicts properly, whether given half a boat or > half a word ex. 't h a n k s g i v ?' it is able to recognize it even if > only saw a small version of a boat or word thanksgiving. This can't be just > delay recognition with discounts for uneven timely activations, it needs a > pattern of error in time errors, Hinton calls it "equivalence". And the > thing is, BACKPROP and whatever cannot "learn" this on their own because it > is too specific a function....it must be in the code, the code is like 6000 > lines of code for heavens sake, which is VerY annoying and I refuse to > understand it for now cuz mine is ~120 lines and half way there. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tb5526c8a9151713b-Mbe7251a66f200e8a3e9fb799 Delivery options: https://agi.topicbox.com/groups/agi/subscription
