Thread by @WiringTheBrain

Thread incoming 🧵 1/13

https://twitter.com/WiringTheBrain/status/1694391227725001061

The concept of “representations” offers a crucial bridge between brain and mind – a way for physical (patterns of neural activity) to manifest as mental; for organisms to be able to *think about things*. 2/13

But representation talk is controversial and laden with baggage. Are they discrete symbolic objects of cognition or distributed states in a dynamic connectionist network? Are they needed at all to explain cognition? How does the meaning of neural patterns get grounded? 3/13

Representations are often referred to as “vehicles with content”. But is that an apt phrasing? Do representations “contain” information about whatever they’re about? How would it get in there? How would the organism make sense of it? 4/13

Of course, there are less literal ways to interpret “content”, but those conceptual connotations come along for the ride and can create confusion and invite controversy. 5/13

An alternative way of framing it is that the meaning of some pattern of neural activity is not *contained within it* but *evoked by it*. (In the same way that the meaning of these words is evoked in your mind, not contained in the squiggles on the screen) 6/13

An evolutionary view shows how internal patterns (biochemical or neural) can become meaningful for an organism, starting with pragmatic control signals and policies, and elaborating to internalised representations that decouple sensation and action. 7/13

These patterns thus start out as meaningful in two ways – they are referential (about something) and – at least potentially – consequential (for something). They’re grounded from the get-go. 8/13

When a hierarchical network of intervening layers evolves, an extra dimension of *relational* meaning can arise – objects have meaning in terms of their known (learned) relations to other objects. 9/13

Note that for artificial systems like Large Language Models, relational meaning is all they have – a self-enclosed world of arbitrary symbols pointing at each other. 10/13

In living organisms, for both pragmatic and semantic patterns, the “content” (or meaning) is really defined by *context* – stored control policies or stored knowledge (schemas). That is, the meaning inheres in how the patterns are interpreted. 11/13

This view offers a way to reconcile computational and connectionist theories of mind. There really are discrete patterns that can act as symbols and be operated over, but the meaning of those symbols inheres in the distributed configuration of the network. 12/13

And that’s all I have to say about that (here). See the preprint for more ... 13/13 psyarxiv.com/dfkrv

@jamessseattle I like Nick Shea's phrasing: the representation "stands in exploitable relation" to some thing or state of affairs in the environment