Monday, 7 October 2019

Show Me the TALoN! (Thoughts on Raja & Anderson, 2019)

There is a special issue of Ecological Psychology out with contributions from lots of people (including us) on what a Gibsonian neuroscience might look like. I'll work my way through the papers over the next few weeks - today, we read Raja & Anderson's contribution for our lab meeting, and I wanted to write about where we ended up with this paper. The upshot is that the paper is clear and the basic ideas of neural reuse and Transiently Assembled Local Neural Subsystems (TALoNs) really do match up nicely to the perception-action scale explanations in the ecological approach. However, it's just not yet clear how much value is added to the ecological approach by these concepts; neural reuse is perhaps not that radical a notion, and there isn't yet any good evidence that TALoNs are a good account of actual neural architecture. As a functional level description of an ecological approach to brains, it seems quite nice, but there isn't anything convincing in here that this is actually how brains work. Show me that, then let's see what happens.

Raja & Anderson propose here that neural reuse (Anderson, 2014) is a good candidate framework for a radical (non-representational) embodied cognitive neuroscience. They note that two basic requirements are that to be usefully Gibsonian, the framework must be non-information processing/non-representational, and that it must click with the behavioural scale theory the ecological approach has now. The paper is a sketch about how neural reuse accomplishes these things.

1. What is neural reuse?

Neural reuse is Anderson's hypothesis that "...different parts of the CNS are used and reused to accomplish different functions at multiple spatial scales...the structure:function relationships in the brain are many:many..." (pg 171). In contrast to a more modular account, this approach suggests that there are no parts of the brain with dedicated functions. Instead, these regions have various capabilities, and express these in softly-assembled task specific coalitions of regions called TALoNS (Transiently Assembled Local Neural Subsystems). 

2. The radical bit (no representations doing information processing)

Raja & Anderson argue this comes pretty naturally to neural reuse. In this account, brain function is never the result of the operation of fixed networks implementing specific processes. Instead, brain function is the result of the higher-order relations between regions (i.e. TALoNs). Regions do have specialisms; not everywhere can do everything. But that specialism is a functional footprint, rather than a computational process.

3. The being-like-our-behavioural-theories bit

Ecological psychologists treat behaviour as an achievement of the organism (and not just of the brain). We have spent our time focused on developing explanations of behaviour grounded at the ecological scale of perceiving and acting. Raja & Anderson want to show that neural reuse can connect to these explanations and be a framework for explaining the contributions to behaviour from the neural scale. 

TALoNs are synergies, literally like the kind of synergies we discuss at the behavioural level (e.g. Latash, 2008). They are composed of pieces, but the abilities of those pieces is the wrong level of analysis. In the same way as 'what the finger does' depends on whether it is in the context of prehension, pointing, or punching, 'what that bit of brain does' depends on the TALoN it is currently part of. The correct unit of analysis is 'the synergy in context' in both cases.

These neural synergies will be constrained by, and also constrain, the behavioural level dynamics. They talk about Fajen & Warren's (2003) model of navigation as an example. In that model, visual information about obstacles, apertures and paths are coupled to bodily dynamics so the latter can be controlled with respect to the former, but 'navigation' is a property of the perceiving-acting agent. In the same way, the relevant TALoNs would involve parts of the brain particularly capable of working with optic flow, and other parts particularly interested in bodily dynamics and coordinating their activity, but because the unit of analysis is the neural synergy in context and not the parts, it would make sense to talk about it as an intrinsically perception-action, embodied contributor to behaviour. 

The pitch here is that all the dynamical systems talk we do about perception-action couplings and synergies works very naturally for TALoNs too, so we can discuss both the ecological and the neural scale in the same language (dynamical systems theory) very naturally. 

Some Thoughts

The basic proposal (that neural reuse and the ecological approach make fairly natural bedfellows) works, I think. But after a spirited lab meeting, I've come away thinking that there just isn't a lot of meat on this paper. 

How radical is neural reuse?
The first issue is that neural reuse isn't necessarily such a radical notion. The idea that different parts of the brain take part in lots of tasks isn't especially controversial, and strong modularity isn't really a feature of modern cognitive neuroscience. Yes, there are commitments that brain regions specialise in terms of the computational processes they implement, but network thinking has pretty thoroughly infected neuroscience and even the idea that these networks are dynamic, not static seems uncontroversial. I do think that Anderson's specific theory emphasises these as key features in a very important way, and that there are important implications from this emphasis. But, for example, a lot of these implications seem like the kind of thing a cognitive neuroscientist could roll with. For example, the claim that 'For neural reuse, the main evolved function of the brain is the control of action...' (pg 171); Daniel Wolpert would agree! 

Anderson's basic evidence for neural reuse are a pair of meta-analyses he performed on neuroimaging studies, in which he found that a) regions of the brain showed up as active in a wide variety of tasks and b) different tasks should entail different combinations of regions. Both are true. However, the tasks were classified based on the cognitive level analysis (so something might show up in both a visual and a language task), and it's been time to relabel the brain for a long time; the 'reuse' might just reflect that the cognitive task analyses are incomplete or incorrect. Second, it seems trivial that different tasks would entail different sets of regions working together; how could it not?

Whither TALoNs?
If there is meat to be had in this proposal, it's to be found with TALoNs. If this is the basic organisational principle of the nervous system, that's pretty interesting and it certainly aligns with the ecological scale. The main problem is that there just isn't much evidence that TALoNs are a thing. Anderson introduces them as a conceptual tool in his book, but I don't know of any experimental work that has found evidence that this tool does better in explaining some neural activity than currently available network concepts. 

I think the idea of TALoNs makes complete sense; I tend to think that there is a lot of converging evidence in favour of these kinds of structures, in perception-action and more complex coordinations across individuals like flocking and herding. But as Anderson himself notes in his book (pg 113), we just don't have to tools to show them empirically yet. He's critiquing standard cognitive neuroscience, but his point applies to TALoNs as well. 


I found myself reading this paper as a fairly sensible but functional level suggestion; good ideas for ecological ways of talking about the brain but nothing very convincing about whether the brain actually is ecological. For example, the paper leaves a key question wide open - how exactly is all this tied together? Raja and Anderson go no further than to say '...these TALoNs are constituted by means of interactive differentiation and active search; namely, they are generated and soft-assembled by a kind of action.' (pg 175). Sure; but this is literally a functional level description of a hypothetical process, with no reference to which pieces of the system do what and when. They've posed a question, they haven't provided an answer. Our Ecological Representations paper in this special issue had taken a more solid swing at thinking about how-actually the brain does what it does, vs just how-possibly it does what it does, and while neural reuse and TALoNs would certainly fit with where we are heading, nothing in our account stands or falls depending on whether these ideas are right or not. 

I don't hate this paper; quite the contrary, I thought it was clear and useful, and the basic idea makes complete sense to me as the right way to go about talking about things. But just as with the free energy principle work, I didn't see anything that added any value to the ecological approach. Show me a TALoN, and then let's talk. I hope I'm wrong :)

No comments:

Post a Comment