On Thu, Jun 30, 2022 at 2:18 PM Boris Kazachenko <[email protected]> wrote:
> On Thursday, June 30, 2022, at 6:10 AM, Rob Freeman wrote: > > what method do you use to do the "connectivity clustering" over it? > > > I design from the scratch, that's the only way to conceptual integrity in > the algorithm: http://www.cognitivealgorithm.info. > I see. So "shared connectivity" not so much in the sense of being connected together. But in the sense of having the same internal connectivity within two groups which are not directly connected together. OK, good. It's good to be clear you are thinking of "connectivity" in another sense. That's a valid sense. I'm focused on the shared observed prediction sense. So your sense didn't occur to me. But that could be a sense of shared connectivity too. So, how about this. Could we use both? If two separate clusters share a prediction, I don't see why you could not then connect them through such shared predictions which occur in a data set, without doing a direct comparison of their respective internal connectivity? You might think of it as somewhat the reverse of what you are doing. I understand you to be comparing connectivity, and predicting based on that. I'm suggesting that if two clusters will tend to share predictions, that might be revealed more directly by such predictions as are observed already. Now, where my idea might come unstuck, is where there are two clusters which might be used to predict the same things on the basis of their internal similarity, as you suggest, but actually they have never been observed to share any predictions. In which case, yes, you would need to directly check any similarity in their clustering, and yours would be the only mechanism. In language that might equate to two words which "mean" the same thing, but which have never been observed to be used in the same sequence. Actually, dredging back... I think we have some evidence for what happens in that case. Taking from what I wrote somewhere else: <<< I think the evidence from language learning is somewhat the opposite. It starts with the particular, and only generalizes later. I always found examples in this study by Peter Howarth some years ago a striking example of this (all appear to be paywalled these days, unfortunately): Phraseology and Second Language Proficiency. Howarth, Peter. Applied Linguistics , v19 n1 p24-44 Mar 1998 ... What interested me was his analysis of two types of collocational disfluencies he characterized as "blends" and "overlaps". By "overlaps" he meant an awkward construction which was nevertheless directly motivated by the existence of an overlapping collocation: e.g. "Those learners usually _pay_ more _efforts_ in adopting a new language..." *pay effort PAY attention/a call MAKE a call/an effort So "*pay efforts" might be motivated by analogy with "pay attention" and "make an effort" (because of the overlapping collocations "pay a call" and "make a call".) In Howarth's words (at least in my pre-print): "Blends, on the other hand, seem to occur among more restricted collocations, where the verbs involved are more obviously figurative or delexical in meaning and the nouns are semantically related, though there are no existing overlapping collocations. '*appropriate _policy_ to be _taken_ with regard to inspections' TAKE steps ADOPT a policy ... It is remarkable, firstly, that NNS writers produce many fewer blends than overlaps and, secondly, that it is the more proficient (by informal assessment) who produce them." What I understand Howarth to be saying is that "overlaps" tend to be produced first, and conceptual "blends" only later. It is the opposite of what we would expect if language learning started by combining words according to general concepts. <<< So on that evidence, yes, the internal similarity mechanism you suggest might be a valid one. It might be used. But the similarity based on observed shared context/prediction mechanism I'm suggesting at least appears to exist, based on what we observe from Howarth's "overlaps", and might be a stronger mechanism. For natural language, anyway. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T5d6fde768988cb74-M8c23a2e9598669e33bc3f173 Delivery options: https://agi.topicbox.com/groups/agi/subscription
