home.social

#cognitivearchitecture — Public Fediverse posts

Live and recent posts from across the Fediverse tagged #cognitivearchitecture, aggregated by home.social.

fetched live
  1. At Thiên Dương Cognitive Architecture Lab, nature is viewed as a data system of the universe. The green crystal is Olivine, formed deep in Earth’s mantle and brought to the surface by volcanic lava. The red crystal often found in volcanic rocks can be Garnet or sometimes Ruby. These minerals record extreme thermodynamic processes where order crystallizes from chaos—an analogy for cognition and information structures emerging from entropy.
    #AI #Geology #EarthScience #CognitiveArchitecture

  2. Idea for AI builders:
    LLMs still behave binary at the edge—fluid → snap.
    Proposing a graded “Breathing Sheath” between emergence and safety:
    a dynamic modulation layer using fractal timing (micro/meso/macro), intention-awareness, and agitation-damping.
    Nuance stays; safety holds.
    Could this be a root architecture for smoother, steerable LLMs?

    #LLM #AISafety #Alignment #AIResearch #ControlTheory #CognitiveArchitecture #AIDev #AIUX #AIFutures #TechFediverse

  3. CW: shop talk

    A neat little (perhaps Soar-specific) aside:

    I find explicit nulls so helpful in our agent's knowledge representations. "I don't think this implies anything meaningful next" as an explicit representation is so much more powerful than relying on the implicit "absense of implication" version. It's the difference between expecting nothing and not knowing what to expect.

    There's a similarly flavored benefit to using ternary logic representations in the agent. Instead of an implicit unknown, you get "I can't evaluate this yet", an explicit signal to gather information.

    It's just nice, that's all.

    #CognitiveArchitecture #AI

  4. I keep thinking about how many tasks are left on our cognitive architecture research to-do list.

    Opportunity cost is real.

    I guess if I had to pick, I'd want to pick embodiment-integrated memory as a reconceptualization of what perception means in the face of one of its functionalities being interfacing with cognition. In other words, semantic memory, perception, and cognition all really using different functional access to an underlying embodiment-world memory/modeling data structure. Think "when tenenbaum was conceptualizing of literally a game engine kind of thing in the head" style "language of thought". I'm not being clear about it, but it's been on my mind awhile.

    A close second is implementing a theoretically complete metacognition for Soar. (Think "sufficient", not "optimal".)

    #AI #CognitiveArchitecture

  5. A lot of my research has essentially been:

    If we add x additional knowledge to the agent about itself, that will naturally afford it y more autonomy capabilities. Ideally alongside a proof by induction style argument about lifelong learnability for that knowledge.

    With the idea being that different types of knowledge you have about yourself might be more task general than explicit knowledge about the world.

    The primitive and to-this-day motivating case being what was already in the system before me: knowing you don't know what to do.

    #AI #CognitiveArchitecture

  6. IMO, universal subgoaling in Soar as an engineering-with-respect-to-its-implementation-of-cognition solution to metacognition is underrated.

    Yeah, Soar doesn't embodiment-representation-learning well yet, but c'mon.

    #AI #CognitiveArchitecture

  7. This week’s book symposium on Movements Of The Mind is now under way!

    You can find all of Dr. Wayne Wu's daily posts here: philosophyofbrains.com/categor

    Wayne (#CarnegieMellonUniversity) plans to focus on comments on each post (rather than on social media).

    Thanks to #OxfordUniversityPress for continuing to partner with us on #book symposia!

    #cognitiveScience #cogSci #philMind #Attention #CognitiveArchitecture #Agency #Bias #Science #PhilosophyOfScience #Neuroscience #Psychology #blog #openAccess

  8. Along with the gactar 0.13.0 release yesterday, I released v0.4.0 of the VS Code extension for it.

    This extension provides syntax highlighting, snippets to quickly create an amod file or add productions, and allows you to run gactar directly through VS Code.

    marketplace.visualstudio.com/i

  9. 🧠 I released v0.13 of my gactar tool for creating & running basic ACT-R models on multiple implementations using a single declarative file format.

    ACT-R (Adaptive Control of Thought — Rational) is both a theory of cognition and an implementation of that theory in software.

    github.com/asmaloney/gactar/re

  10. "types of knowledge" in cognitive architectures is one of those things that can either be

    - very helpful taxonamization by necessarily entailed computational properties or
    - putting the cart before the horse and commiting to meaningless divisions up front that muddle or remove the generality of the architecture.

    It's a hard thing to navigate. Think about the "camps" of various hypotheses in neuroscience and how hard it is to form a consensus on (for example) the real "contents" and "function" of the entorhinal cortex w.r.t. semantic memory, episodic memory, and perception.

    #CognitiveAI #AI #CognitiveArchitecture

  11. This is actually a great Friday task - updating documentation for our procedural memory learning system.

    Time to crank up the music and get it done.

    #computerScience #cognitiveArchitecture

  12. 🧠 I released v0.12 of my gactar tool for creating & running basic ACT-R models on multiple implementations using a single declarative file format.

    ACT-R (Adaptive Control of Thought — Rational) is both a theory of cognition and an implementation of that theory in software.

    This release fixes issues, adds some capabilities to the language, & adds some new command-line tools.

    github.com/asmaloney/gactar/re

  13. @pinecone If you appreciate a systems approach, you may love #CognitiveArchitecture research. Here is one based on Allen Newell's argument that cognition is a systems problem.
    Soar: soar.eecs.umich.edu/

    We even studied language + situated action before it was #AI #ML cool.

    Comprehension: arxiv.org/pdf/1604.02509.pdf
    Learning grounded language: cogsys.org/journal/volume2/art
    Instruction: arxiv.org/abs/1604.06849

    And yes, knowledge is reasoning about operating in the world.

  14. For any cognitive science people out there in the Mastiverse who are into *symbolic*, I've just updated my mirror of Lisp ACT-R on GitHub to **7.27.7 **from 19 July 2022.

    github.com/asmaloney/ACT-R

    (For non-CogSci people, ACT-R (Adaptive Control of Thought — Rational) is both a theory of cognition and an implementation of that theory in software.)

  15. 🧠 I released v0.11 of my gactar tool for creating & running basic ACT-R models on multiple implementations using a single declarative file format.

    ACT-R (Adaptive Control of Thought — Rational) is both a theory of cognition and an implementation of that theory in software.

    This release changes it to a command-style cli & adds colour to the command-line output.

    github.com/asmaloney/gactar

  16. @AbelTM @garymarcus I can see that. I don't know Bengio's approach in full detail.
    But, Laird's approach has been that of integration - as computational systems mature, he attempts to bring them into the #CognitiveArchitecture

    For instance, we are current studying what it means to connect (theoretically and systematically) #CV #transformer models in visual reasoning with #cognitive #relational #representations similar to what a #AIPlanner would use.

  17. Laird is from the #AllenNewell school of #UnifiedTheoriesOfCognition
    en.wikipedia.org/wiki/Unified_

    Allen famously stated "You can't play 20 questions with nature and win!".
    mindhacks.com/2015/02/10/you-c. We can't implement cognitive capabilities - vision, NLP, planning, RL - separately and claim to understand human intelligence. Human brain does all of this in a single #architecture.

    #Soar is Laird's approach to that challenge - how do we build a #CognitiveArchitecture
    soar.eecs.umich.edu/

    #AI #ML

  18. This weekend will be exciting! John E. Laird (laird.engin.umich.edu/), Yoshua Bengio (yoshuabengio.org/) and @garymarcus are speaking at #AAAI #FSS 2022 on Thinking Fast and Slow - sites.google.com/view/aaai-fss.

    Laird and Bengio represent the cognitive-symbolic and statistical-sub-symbolic views respectively. Both have architectures that implement some form of 'fast and slow'. @garymarcus ofcourse is the chief cynic :)

    #AI #ML #ThinkingFastAndSlow #CognitiveArchitecture #NeuroSymbolic

  19. [Thread (1/n)]: In the #paper "A #Storytelling Robot managing Persuasive and Ethical Stances via ACT-R" (available here: arxiv.org/pdf/2107.12845.pdf) and published on the International Journal of Social #Robotics, we present a #persuasiverobot integrating (via the ACT-R #cognitivearchitecture) #storytelling strategies, #ethical stances, and #rhetoricaltechniques (based on #framing and on the use of some fallacious arguments like the "argumentum ad verecundiam" and the "argumentum ad populum"). 👇

  20. @davidbeers

    You may also find the Common Model of Cognition paper - ojs.aaai.org/index.php/aimagaz by Laird et al. also interesting.

    It is not modern per se because the #CognitiveArchitecture research is 40 years old at this point. But, we are still advancing it as new computational methods are discovered.