I know it doesn't matter, but I am so confused by this idea floating around that a dynamical systems approach is somehow without representations. It is not how I nor many other computational neuroscientists I know would think of these things. www.cambridge.org/core/books/r... - ThreadSky

neurograce.bsky.social • 37 days ago

I know it doesn't matter, but I am so confused by this idea floating around that a dynamical systems approach is somehow without representations. It is not how I nor many other computational neuroscientists I know would think of these things. https://www.cambridge.org/core/books/representation-reconsidered/where-is-the-representational-paradigm-headed/7220BDDFB1FA2C0B833C3A16D66956D7

Comments

paultopping.bsky.social•37 days ago

Your quote says to me that they are not talking about systems that contain no representations at all but that they don't explain their inner workings in terms of representations. It's as if they're saying, "We don't know how they do what they do and, quite frankly, who cares?"

jbrendanritchie.bsky.social•37 days ago

This was more a thing in the 90s tbf, popular in phil. Eliasmith wrote some papers then on why reps are compatible with DST, also Toribio + Clark 94. More recent folks mostly take the same line unless they're going more DST + 4E COGNITION, eg Chemero. But the maths math whatever the theory.

romanfeiman.bsky.social•36 days ago

I think this paper by Tim van Gelder gives a good intuition for what an alternative to representations in this framework might be. There’s a giant promissory note in the main analogy, but at least it’s one worked example. https://www.jstor.org/stable/2941061

romanfeiman.bsky.social•36 days ago

Braitenberg vehicles are another fun example. Again, unclear how far the analogy could really extend, but the eliminativist intuition about what these vehicles do and don’t represent doesn’t seem confused to me.

romanfeiman.bsky.social•36 days ago

https://www.usna.edu/Users/cs/crabbe/SI475/current/vehicles.pdf

kihbernetics.bsky.social•34 days ago

IMHO, representations, computation, and modeling are all necessary only in #anticipatory dynamical systems that can make #predictions using current #knowledge to attain #information by comparing the #expected state of the environment with the one that was #perceived through sensory "infrastructure."

spiantado.bsky.social•36 days ago

It's also funny that many don't seem to know that there are discrete dynamical systems (including, um, Turing machines)

manolomartinez.net•36 days ago

I think they mean: if we are going to provide simple descriptions of some cognitive function, sometimes a few ODEs do a better job than an algorithm over symbol-like entities.

withwirth.bsky.social•37 days ago

I work from and have been trained in the approach that is being talked about here, and I think the idea that dynamic systems are without representations is confused-it entirely depends on what the dynamics are describing.

Also how you define representation (as others point out)

johanneskleiner.bsky.social•29 days ago

📌

steveread.bsky.social•36 days ago

I think for many people their intro to dynamical systems thinking was in terms of physical systems like the weather or systems with a single variable where representations didn’t play a role

jonathanamichaels.bsky.social•37 days ago

Such a weird take. I feel like it's a hallucinated controversy chatGPT would make up

hritz.bsky.social•36 days ago

jonathanamichaels.bsky.social•36 days ago

Trolling me with my own paper 😅

hritz.bsky.social•36 days ago

🫢

cmcshaff.bsky.social•36 days ago

As many are saying, it depends a lot on personal definitions of those words. As someone from a lab where nonlinear DST is used to study small neural circuits in simulated agents, I will say that a lot of us have had our notions of representation thoroughly challenged while walking through the math

mitchherschbach.bsky.social•37 days ago

One reason could be most of the advocates of this view in the 1990s weren't neuroscientists--e.g., Tim van Gelder, Randall Beer, Esther Thelen, Linda Smith. They were cognitive scientists defending embodied cognition or extended cognition. Section 4.2 of this article covers some of this literature

haikugoat.bsky.social•37 days ago

Do you by any chance know of a good survey/intro article on this dynamical systems way of looking at things? Hopefully something not too technical? Hopefully something that situates it in relation to connectionism? Thanks! My background is in formal linguistics mainly.

haikugoat.bsky.social•37 days ago

But when I was a cig sci undergrad at Johns Hopkins in early 2010s, there seemed to already be an emerging consensus amongst the faculty that the apparent tension b/w symbolic representations and neural networks was illusory...

haikugoat.bsky.social•37 days ago

I think that sentiment played out in a very connectionist setting (Paul Smolensky was a big presence). I'm trying to understand how dynamical systems approaches fit into that whole dialectic

ebben.bsky.social•36 days ago

Also, "representation" has to include action patterns. An action pattern, like a horse's various gaits, are representations, in this case of how to move in terrain. This theoretical aspect of memory over time ("abilities") is probably the most contentious issue and part of the confusion.

ebben.bsky.social•36 days ago

I think the confusion stems from old views of symbolic processing: "cognitive processes are the rule governed manipulation of internal symbolic representations"

The term "representation" has a much broader, non-symbolic meaning now, and thus not necessarily in conflict with dynamical theory.

jbarbosa.org•37 days ago

I agree with you that this particular wording is very weird, but I do see a tension in the field between those that look primarily at representations and ignore dynamics and vice versa. Again, this is not what the author seems to be alluding to.

lincoln81.bsky.social•35 days ago

In your view are representations doing causal work? Or are they part of an explanatory story, but can be jettisoned by a causal story once we have the underlying math / neurobiology correct? Because if that’s the case then I don’t think there’s a conflict.

neurograce.bsky.social•35 days ago

I think I understand what you're asking and I would maybe say yes. But really I don't think that you jettison the representations once you know the dynamical systems because they're just (semi-)different ways of describing the same thing. It only depends on what your research question is.

lincoln81.bsky.social•35 days ago

Yes, well I think that is close to the position that’s described in the quote. The contrast would be something like that you can’t give a causal explanation without representations. They’re not merely different ways of talking about the same thing.

neurograce.bsky.social•35 days ago

Does that mean that representations are thought of not as something that is implemented by a physical system?

lincoln81.bsky.social•35 days ago

No, the vehicles that carry the representations are physical things

neurograce.bsky.social•35 days ago

So then couldn't those physical things be described in terms of their physics (a more dynamical systems approach) or the representations they carry?

neurograce.bsky.social•35 days ago

I think this is partly why people were opposed to the Blue Brain approach of doing whole brain simulation. It felt like describing the dynamical system without talking about what it did or why that matters.

theresecollins.bsky.social•37 days ago

I gather from the responses that this view is not shared by yawl computationalists. But can someone explain why someone might think dynamical systems approach is without representations?

dryohanjohn.bsky.social•35 days ago

Yup! I have come across this several times with philosophers. Very odd indeed.

The historical key to understanding this is that Jerry Fodor convinced a lot of people that the only representations that count as representations are symbolic representations.

pessoabrain.bsky.social•37 days ago

No offense to either side but this could be a divide between "older" and "younger" investigators and how this debate was cast in the 1980-90s. The 1970-80s was a time of symbolic computation with production systems, etc. Then PDP models popularized neural networks and dynamical systems.

jpober.bsky.social•36 days ago

Try explaining this to people at a 4E conference

(I did once, not sure I plan on doing so again 😅)

pessoabrain.bsky.social•35 days ago

Scientists need to be a bit more open minds imo...

mraginsky.bsky.social•36 days ago

Right, one has to understand "representations" in the Fodorian (or even GOFAI) sense as logical propositions related to the state of external environment via a sequence of truth-preserving transformations. Vectors of activations in a connectionist system are not representations in this narrow sense.

mraginsky.bsky.social•36 days ago

But one can embed symbolic logic in either discrete or continuous dynamical system, so there is really no opposition here. Most of the invocations of dynamical systems in cog-sci or even neuro are hopelessly confused, a sad state of affairs brought about by lack of dialogue with control theorists ;)

mraginsky.bsky.social•36 days ago

With very few exceptions, the mathematics of dynamical systems in cog-sci or in neuro has been taken from physics, where one typically considers closed systems without external inputs or outputs. I wrote about this in some detail here: https://link.springer.com/article/10.1007/s13752-023-00440-6

neurostats.org•30 days ago

📌

neurostats.org•30 days ago

Ah! Now I understand why I just couldn't follow what the cog-sci folks were doing!

pessoabrain.bsky.social•30 days ago

It was a wild idea, not completely dead though

kihbernetics.bsky.social•34 days ago

Maturana and Varela's #autopoietic (enactivist) view of organizational / operational closure is compatible with the traditional cybernetic paradigm based on inputs, outputs, feedback, and internal (state) representations when seen as the #allopoietic "outer layer" sustaining the structural coupling.

hritz.bsky.social•36 days ago

love this essay!

mraginsky.bsky.social•36 days ago

Thank you!

cmooreanderson.bsky.social•36 days ago

Could you send me a copy of this, please?

justfineneuro.bsky.social•37 days ago

And on representations and dynamics, let’s take a classic eco psych topic: optical flow and locomitio. If a network codes the lawful transitions between optical states, and self motion is coded, then network can code for the dynamics. Does it not represent in classic sense using neural integration?

justfineneuro.bsky.social•37 days ago

There’s a weird history in psychology dealing with this, starting with people in ecological psych fields (Gibsonian). And they were desperate to be as non-representational as possible, were really the first to use dyanamics (Turvey, Kelso). That articles date tracks with that insular thinking.

adw.bsky.social•37 days ago

“Desperate to be as non-representational as possible” is a weird framing. They have reason to think representations were an error and dynamics offers a way to do things without them.

wrt the OP you can do dynamical representations, it’s just not compulsory

justfineneuro.bsky.social•37 days ago

It wasn’t meant to be inflammatory. More a reflection of the training I had in grad school, where discussion of representational building blocks was scoffed at rather than treated with intrigue.

adw.bsky.social•37 days ago

No one thinks the atmosphere represents or computes things; it’s structure results from nonlinear self organisation under constraints. The dynamical hypothesis is that this is the right way to talk about cognition too

kordinglab.bsky.social•36 days ago

Mark Churchland very much promoted this idea (eg in the jpca nature paper). I think it helped popularize the "dynamics" idea to phrase it as an attack on a strawman.

shahabbakht.bsky.social•37 days ago

This has always bugged me too, but I think it’s likely caused by the incoherent or ambiguous use of the terms involved, as is often the case with persistent disagreements. Like, I’m sure people mean different things when they use "representation."

neurostats.org•30 days ago

Yes I think the representationalism in phil.neuro is sort of distinct from what practical neuroscientists were doing. One was making certain theoretical commitments but the other was simply methodological representationalism. But somehow all these things were so mixed up, esp. in 2011-2015

shahabbakht.bsky.social•30 days ago

Exactly. And depending on their perspective, people often take away messages that go beyond what the authors or the paper even intended.

neurograce.bsky.social•37 days ago

And "dynamical systems"

tyrellturing.bsky.social•37 days ago

At the risk of being a jerk, I think this is bc many of the people originally promoting "dynamical systems theory" were philosophers who didn't really understand the math of dynamical systems or ANNs.

So, they didn't realize the mathematical guts of the two approaches are essentially identical.

pnwnet.bsky.social•36 days ago

"dynamical systems theory" sounds like "complexity theory" which is a branch of thinking different from reductionist approaches, and would better describe how nature works based on complex systems and not parts.

gutlin.bsky.social•36 days ago

Haha true. Really weird.

You can literally just take a Deep Neural Network, move it from the discrete to the continuous domain, and voilà: you have a perfectly fine "dynamical system" LIF Spiking Neural Network (the representations are still there. SNN people just don't focus on them that much)...

gutlin.bsky.social•36 days ago

Like, they are perfectly equivalent under certain circumstances:

https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2022.857513/full

tyrellturing.bsky.social•37 days ago

But, a more charitable interpretation is that many of them also believed/hoped that one could develop a dynamical systems theory of behaviour absent any reference to internal workings of the brain. In which case, you are perhaps not positing representations in the brain itself in your model.

hritz.bsky.social•36 days ago

Ya, remember Eliasmith was skeptical of this aspect of dynamism.

good paper iirc
https://www.researchgate.net/profile/Chris-Eliasmith/publication/28762416_The_third_contender_A_critical_examination_of_the_Dynamicist_theory_of_cognition/links/545abbc00cf25c508c319aaf/The-third-contender-A-critical-examination-of-the-Dynamicist-theory-of-cognition.pdf

neuroai.bsky.social•36 days ago

His own “alternative” approach is called representation and dynamics.

mraginsky.bsky.social•36 days ago

https://bsky.app/profile/mraginsky.bsky.social/post/3lg7xiwphls2i

neuroai.bsky.social•36 days ago

Yup https://bsky.app/profile/neuroai.bsky.social/post/3lg7x7o7s4s2e

neuroai.bsky.social•36 days ago

Agreed. Plus another issue: the philosophical notions of representation in that work are bizarrely strict. When I was a neuro grad student I remember a bunch of philosophy grad students laughing at me at a conference for saying I was studying neural representations.

tyrellturing.bsky.social•36 days ago

Yeah, I fully agree, that's the other part of it: they employ such an overwraught definition of "representation" that they can claim to be advancing a theory sans representations even when the models contain what the vast majority of scientists would call representations.

nmwilkinson.bsky.social•36 days ago

Not sure about this particular example in the OP quote, but in some cases it was a reaction to what was (I think rightly) perceived as a situation where representations had become the "billiard balls of cognitive science" (Inman Harvey). Whenever some e.g. stimulus distinction was observed, it...

nmwilkinson.bsky.social•36 days ago

*stimulus X

...was taken to imply "internal representations of X". This would just be asserted without further evidence. Example from my own work - newborns (A) seem to be able to distinguish "facelike" stimuli, therefore (B) they must have an innate representation of faces. But turns out...

nmwilkinson.bsky.social•36 days ago

...a does not imply B at all, as we showed here👇. Not the same as the idea that nothing one might classify as a representation exists in the system, but this was a very standard misinterpretation with wide-ranging implications (and is often still seen in psychology)
https://doi.org/10.1111/desc.12159

neuroai.bsky.social•36 days ago

tyrellturing.bsky.social•36 days ago

richarddmorey.bsky.social•36 days ago

Maybe I'm misreading, but isn't the key part "as an explanatory posit"? As in: you can use representations descriptively, but they aren't key to *explanations* per se.

keithfrankish.bsky.social•36 days ago

Yes, Chemero makes this point forcefully in REC.

davidfarrokh.bsky.social•35 days ago

What do you mean by representation? The study below shows the degree of confusion in applying the term.

Hard to take a strong stance without knowing what people mean, although the general confusion surrounding the term is quite telling, imo.

https://www.frontiersin.org/journals/psychology/articles/10.3389/fpsyg.2023.1165622/full

jonathannajenson.bsky.social•36 days ago

I think the motivation was (is?) that dynamical systems need not entail representational talk, unlike classical computation. Totally agree that they work well together though

jonathannajenson.bsky.social•36 days ago

Randy Beer has a nice piece on this -
https://onlinelibrary.wiley.com/doi/full/10.1111/cogs.12142

andrewmurphie.bsky.social•37 days ago

I think the confusion is often between representation vs non-representation on the one hand, and, on the other, dominance of representational processes vs primary importance of other processes. This before you get to rep/other process combinations. Lots of tensions on this in humanities too.

manuelbaltieri.bsky.social•36 days ago

One thing is to assume there must be a “representation” and use DST to study its dynamics, one thing is to couple two open dynamical systems and give necessary/sufficient conditions to explain (non-trivially) when one “represents” the other.

manuelbaltieri.bsky.social•36 days ago

The latter is far from trivial and is arguably still a debate to be had (what mathematical proof is there?). It is also a core part of the “alternative” this community proposes, though, admittedly, often not very clearly explained.

beakr.bsky.social•37 days ago

How do you think about this? Can you give an example of a dynamical system explanation where representations play an explanatory role? (I have no view to defend here - just curious)

neurograce.bsky.social•37 days ago

From my perspective, many dynamical systems models are trying to explain how representations arise based on anatomy and neurophysiology.

neurograce.bsky.social•37 days ago

Eg this is work from postdoc where we build recurrent neural networks and try to understand how their representations change over time in order to enhance performance https://www.biorxiv.org/content/10.1101/2022.03.07.483196v1

khakhalin.bsky.social•37 days ago

Hopfield network as the Last Common Ancestor of both Dynamic-non-Representationalists and Network-Connectivists

Comments

Posting Rules

Reply