[FRIAM] Selective cultural processes generate adaptive heuristics
David Eric Smith
desmith at santafe.edu
Fri Apr 15 20:14:04 EDT 2022
Great resource, Glen, thank you,
Your summary is indeed just the point I was after.
Also good to know you have worked in this space.
And the PaLM looks like a project it would be fun to be on the inside of.
I realize, in trying to think what to answer, that there is a(n obvious) meta-question behind the way I asked these things. When is one thing a model for another, in some way that confers insight?
I have, off and on, followed some of the communication-bot projects that come up. The results often look to me like small algorithms running preset courses, with very little they can generate on their own that is a model for anything we have had real trouble understanding about the experience of existing or the resources that language gives us for representing or expressing the nature of that experience. I have this optimism of the amateur outsider that each new generation of computational tasks will have enough richness that the problems we can push them to solve with blunt reinforcement rewards will start to overlap better with the particulars of discourse.
Many thanks,
Eric
> On Apr 14, 2022, at 9:33 PM, glen <gepropella at gmail.com> wrote:
>
> Apologies for the violent reduction of your question. But I'm not smart enough to think with so many words. So I have to reduce it in order to think about it. 8^D Here's my reduction:
>
> 1a Will bots represent their insides?
> 2a Would the (social) coordination of bots be linguistic?
>
> Has anyone tried building a system of Communicating Bots?
>
> I can answer the last with "yes, I have", but not in any sort of academic, scholarly, industrial, or any other serious context. So, really my "yes" is a "no, I haven't". But the project I was on did get at the question of what such things might tell us. I'll skip to the end and render my opinion and only go into the details if needed. We built 2 bots, both intended to interact with children in an augmented reality art display. The bots could talk to both the children and each other. To see how it would go before the event, we set them loose in an IRC chat room. There were 2 notable effects: 1) their conversations usually had a relatively short transient, maybe 100 sentences? I forget. Sometimes it seemed to go on for a long time. But mostly they were short interactions. And 2) they tended to explore the edge cases of their corpora. Each bot was programmed with a different corpus and their interaction tended to the corpus "leaves".
>
> Now, our AI was not really AI as we're talking about it, here. It wasn't ML. We programmed in, manually, an ontology and a grammar by which that ontology was walked. So, whether it would do the same with one of these inductive MLs is unlikely. But, to target your sentiment that it seems like the first thing someone would do, it *was* the first thing we did. 8^D
>
> Also, re the complexity of the gen-phen map, in context learning, etc., this came across my desk recently:
>
> Pathways Language Model (PaLM): Scaling to 540 Billion Parameters for Breakthrough Performance
> https://linkprotect.cudasvc.com/url?a=https%3a%2f%2fai.googleblog.com%2f2022%2f04%2fpathways-language-model-palm-scaling-to.html%3fm%3d1&c=E,1,c8E6lUAGCKhk8Yjf3gSMmVDygrPBDHb3EIPADLxAabc9zjDgQOw7b-osl4MdPwwky20YdrMJ5GnfCTgfSbDmksLXOvbUX98JjSLcyHL85Drj&typo=1
>
> I hadn't read it till my insomnia this morning. But it seems to partially answer the robustness/polyphenism question I asked.
>
>
>
> On 4/13/22 13:35, David Eric Smith wrote:
>> So Glen’s line of questioning here prompts a question in me. Partly ignorant (can’t be helped) and partly lazy (hasn’t been helped).
>> We have two things we have said at various time we would like to understand:
>> 1. Unpacking the black box of whatever-NN representations of patterns;
>> 2. Getting to a “theory” of “what language is” and how we should think of its tokens and structures and acts in relation to not-clearly-conceived notions of “meaning” or “reality”, on which language is supposed to be some sort of window, however dirty.
>> So if we gave a bunch of autonomously-hosted NNs some load of work to keep them all busy, and offered them signal-exchange that, if they happened to, they could find ways to employ to coordinate each other’s autonomously-generated operations:
>> 1A. Would one of the things their emergent signaling systems do be to constitute representations of the “inside” of the “black box” we have been saying we want representations of, not simply coextensive with reporting the input-output pairs that the black box is producing? and;
>> 2A. Would their whole coordination system be in any interesting sense an instance of “language”, and since we would be able to look at the whole thing, and not be restricted to operating through the channel of exchanged signals, would there be anything interesting to learn about the inherent distortions or artifacts or lacunae of language, as referred to some more broadly-anchored senses of “meaning” or “reality”?
>> I could imagine that we would complain that the coordination-traffic purported to be representations of the black box, but that they did a terrible job of actually being that (or one that we couldn’t use to satisfy what _we_ want from a representation), but perhaps we could see some familiar patterns in the disappointments (?).
>> To cut to a small, concrete dataset, we could try it with a kind of bootstrapping partition of the zero-shot translation, in which we autonomously host translation-learning on subsets of languages, but we keep firewalls between the learners so that for every learner, there are languages or language pairs to which it is not given direct access. We have reason to think that each learner would develop a decent version of some internal meta-language, and that there would be coherences of structure across them, since that is what the zero-shot team claims already to have demonstrated. But with different subsets of languages as inputs (and perhaps not less, accidents of the training path just due to noise), there should also be differences and things each one misses. Cross-platform signaling could at least in principle have some available information to convey, as well as much information that it would not need to convey because the shared ancestry of the learning algorithms on each platform makes talking about it unnecessary. The target we would be after in posing the problem is, to what degree is the cross-platform communication insightful to us about the thing we have said we want a representation of (the “meta-language” “within” each zero-shot learner, which is its version of some patterns in the world).
>> Has a lot of this already been done by somebody? It seems like the first thing that somebody who knows nothing would propose to do, so I assume it has already been done to the point where people lost interest and went on to something else.
>
> --
> Mɥǝu ǝlǝdɥɐuʇs ɟᴉƃɥʇ' ʇɥǝ ƃɹɐss snɟɟǝɹs˙
>
> .-- .- -. - / .- -.-. - .. --- -. ..--.. / -.-. --- -. .--- ..- --. .- - .
> FRIAM Applied Complexity Group listserv
> Zoom Fridays 9:30a-12p Mtn UTC-6 https://linkprotect.cudasvc.com/url?a=https%3a%2f%2f%2f%2fbit.ly%2fvirtualfriam&c=E,1,RM7o6LI41b07O-GtvvgCzthiYrldlcdxVwh-lFbAHXoRe_tSY0xaYhxb2ljkAUl_wueX2qE6cQevn8h8pjaGk7S_zn3d7Yz61H-Yc90yz6FlqI48QvHEDr0,&typo=1
> un/subscribe https://linkprotect.cudasvc.com/url?a=http%3a%2f%2fredfish.com%2fmailman%2flistinfo%2ffriam_redfish.com&c=E,1,e15F6qvW82TSHCZBUQp8Bsw9qiAJaYv4mi-RI7qr19gFmXVhoCgPfey_ibtCW_lKtmM2rxzuuA7teMxJgjlDua5JEob3dpBtWjIBbDx8Kn68NcM6_BOZfWEzFT2c&typo=1
> FRIAM-COMIC https://linkprotect.cudasvc.com/url?a=http%3a%2f%2ffriam-comic.blogspot.com%2f&c=E,1,ESyfNx2jvZOYcM4NyHvaTChb9MzsqhNGSe4ZJfObmyi1QykkvLL423xyFUU6czqf0HB24IOL6jCsWDhCXkvo0qEuKKlwanVnp_luPRYvTNnblBu0LBE,&typo=1
> archives:
> 5/2017 thru present https://linkprotect.cudasvc.com/url?a=https%3a%2f%2fredfish.com%2fpipermail%2ffriam_redfish.com%2f&c=E,1,l17ymm-LJ-d8hf-uN266y-R1E3CTB02gRNBSRZ4MP5mRk-7MOKs-Eienb2EGV5BDAhlMG7VMkYCKcEcDprAGIMbkr8y7-RvCdiopQrKAkoXRI3aqpjSW_Hnkg1RN&typo=1
> 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
More information about the Friam
mailing list