I'll try to give some constructive thoughts. In my view I see you talking way too much about the A>B rule, and using too many names to say the same thing, from what I know this is just simply a simple pattern that everything is based on (ex. word2vec...etc). You even made a diagram, it seems overdone on trying to look formal, which wasted time and energy.
"The gist of our theory is that Deep Learning provides us with neural networks (ie. non-linear functions) that serve as the proof mechanism of logic via the Curry-Howard isomorphism. With this interpretation, we can impose the mathematical structure of logic (such as symmetries) onto neural networks." In case not, I do hope you know how my algorithm works. https://encode.su/threads/3595-Star-Engine-AI-data-compressor I don't see the need to use any old fashion logo forumlation or the need to suggest it can be made to work with deep nets which is the same thing as well but wearing a different top (implementation method). Maybe you mean make the implementation of logic combine with the implementation of deep learning, i.e. backprop + very-hardcoded rules. If so, I think what you really want is my algorithm which is just is logic in clean form really and no blurry backprop in the way. Do note mine is not some hardcoded chatbot. My AI is an ultra-advanced markov chain, and I only just begin. The mechanisms are clear in my explanation above and the implementation - as you know - is the code that runs it however you code that thing up to work it fast and efficiently - and others use backprop etc, I do it another way. But the AI is always the same really, there's only one way AGI works, many ways to code it, and one way you should code it. https://en.wikipedia.org/wiki/Curry%E2%80%93Howard_correspondence This too seems wasteful to abstract, logic is algorithm and algorithm is AI, the whole universe is. AI is simply the most common patterns and then it comes up with small mental programs/code (memories) all on its own. See how it mentions in that table truth, sums, categories, implication....these are all explained in my AGI guide I once tried showing to some select few friends. Specifically the truth and sums are not actually coded in my AI and probably won't need to either, it is more rarer needed to do those predictions. "In particular, logic propositions in a conjunction (such as A ∧ B) are commutative, ie. invariant under permutations, which is a “symmetry” of logic. This symmetry essentially decomposes a logic “state” into a set of propositions, and seems to be a fundamental feature of most logics known to humans. Imposing this symmetry on neural networks gives rise to symmetric neural networks, which can be easily implemented thanks to separate research on the latter topic. This is discussed in §3." Again (I believe) I see you doing it here too. It looks like you are trying to hard to abstract it and connect things. "As an aside, the Curry-Howard isomorphism also establishes connections to diverse disciplines. Whenever there is a space of elements and some operations over them, there is a chance that it has an underlying “logic” to it."" All of AI has an underlying logic to do it...all of AI is just built up from markov chain rule. The first pattern you can only find in a dataset is how many times a letter or word repeats, and what follows around it ex. zb or bz or bzq. "Why BERT is a Logic In the following diagram 2, observe that the Transformer is permutation-invariant (or more precisely, equivariant). That is to say, for example, if input #1 and #2 are swapped, then output #1 and #2 would also be swapped:" Happy to see this here. In my plans for my AI I will turn the delay matching into a as Hinton calls it "equivalence" so that h e l l o matches hello, the input has many spaces but most are ignored as they are 'all' spaced, so error is not as big as it would normally think, it matches less only if there is no pattern and much change. As well I know how to teach it abcdefg and show it gfedc and predict the rest backwards. The idea is it matches a lot to the non backwards memory and then instead of predicting the tail next letter it predicts the delay order as the input is seen, it predicts the rest of the memory. And I don't think humans are good at this pattern, we can only do it by hand using lots of resources stressfully. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Tb5526c8a9151713b-M08c70390ee2b0e26d3735eb5 Delivery options: https://agi.topicbox.com/groups/agi/subscription
