Ulrike Hahn's status

Ulrike Hahn replied to Dimitri Coelho Mollo's status

Dimitri, my thinking on the recursion and the context vector was this: yes, it's the context vector input that is causally efficacious in first instance. But I think just focussing on that might miss what the context vector does: in the context of art. neural networks, it makes the system sensitive to a whole new dimension -time- and statistical properties of the environment inherent in the time dimension.

1/2

@dcm @uh @SylviaFysica

Dimitri, my thinking on the recursion and the context vector was this: yes, it's the context vector input that is causally efficacious in first instance. But I think just focussing on that might miss what the context vector does: in the context of art. neural networks, it makes the system sensitive to a whole new dimension -time- and statistical properties of the environment inherent in the time dimension.

1/2

Ulrike Hahn replied to Ulrike Hahn's status

@dcm @uh @SylviaFysica

2/2 this means also that "patterns' have a whole new route for potential efficacy/realism.

Again, it's not particularly well-articulated in the book, but I think there might genuinely be something there.

Imagine a NN that, due to context vectors, becomes sensitive to the possibility of long range dependencies in language.

That is starting to feel like expanding the 'rule set' in some way, and making patterns more real

@dcm @uh @SylviaFysica

2/2 this means also that "patterns' have a whole new route for potential efficacy/realism.

Again, it's not particularly well-articulated in the book, but I think there might genuinely be something there.

Imagine a NN that, due to context vectors, becomes sensitive to the possibility of long range dependencies in language.

That is starting to feel like expanding the 'rule set' in some way, and making patterns more real

Dimitri Coelho Mollo replied to Ulrike Hahn's status

@UlrikeHahn @uh @SylviaFysica I agree, my point was just that, from the point of view of current phil of science, there's nothing very controversial there in terms of ontology, as J. wants to make it. No strong emergence or some such, no top down causation, etc.
The causal sensitivity you mention seems capturable by notions like structuring causes.

Ulrike Hahn replied to Dimitri Coelho Mollo's status

@dcm @uh @SylviaFysica no idea what J. thinks, but my GoL take was that patterns are, in effect, epiphenomenal vis a vis the fundamental layer (and all ‚patterns‘ are equally real…). That intuition hinges for me on the lack of ‚rules‘ in the system that attach to those patterns. So the possibility I just described with recursion and the causal impact of the statistics of the environment seems a new conceptual step beyond the basic GoL model. 1/2

Ulrike Hahn replied to Ulrike Hahn's status

@dcm @uh @SylviaFysica 2/2 it might well be that such effects are well-described as structuring caused. But there‘s a ‚gap‘ in my metaphysical/ontological intuitions whereby I know them to be explanatorily indispensable for explaining behaviour, but have no deep view on how ‚real‘ they are or how they fit into a metaphysical picture (could be: as real as other causes, less real, it‘s all epiphenomenal bar the fundamental level GoL style)….

Dimitri Coelho Mollo replied to Ulrike Hahn's status

@UlrikeHahn @uh @SylviaFysica I don't think those things need go that way. One can recognise patterns and non-fundamental kinds as real and causally efficacious (tables, organisms, trains, etc) and think that they are realized without remainder by lower-level stuff. Only very strong versions of reductionism, currently rather unpopular, would deny that.

Ulrike Hahn replied to Dimitri Coelho Mollo's status

@dcm @uh @SylviaFysica is saying they are „realised without remainder by lower level stuff“ the same as saying that ‘system rules’ at the fundamental level suffice? I thought it is the opposite: to be non-reductionist is to believe that higher level laws do not follow from component parts and lower level laws. But then I have both the problem of how higher level laws are even possible and what impact they can have on lower level ‘stuff’, no?

Dimitri Coelho Mollo replied to Ulrike Hahn's status

@UlrikeHahn @uh @SylviaFysica it just means that the high level laws are realised by lower level ones, but there is no type-type mapping. So high level laws cannot be reduced to lower level ones, but the former are token realised by complex conjunctions of the latter. That does mean though that there can be no top-down causation.

Ulrike Hahn replied to Dimitri Coelho Mollo's status

@dcm @uh @SylviaFysica

so the big question is what gives rise to the high-level laws as types, right? This is where the notion of constraint seems potentially essential and missing from that standard non-reductionist picture since constraints are (metaphysically) neither 'objects' nor laws ('rules')?

and we would want to distinguish as different two metaphysical pictures: one where constraints constrain the fundamental level objects and another where the constrain higher level types qua type?

@dcm @uh @SylviaFysica

so the big question is what gives rise to the high-level laws as types, right? This is where the notion of constraint seems potentially essential and missing from that standard non-reductionist picture since constraints are (metaphysically) neither 'objects' nor laws ('rules')?

and we would want to distinguish as different two metaphysical pictures: one where constraints constrain the fundamental level objects and another where the constrain higher level types qua type?

Dimitri Coelho Mollo replied to Ulrike Hahn's status

@UlrikeHahn @uh @SylviaFysica yes, that sounds like a great way of putting it. In my neck of the woods, the high level types are typically taken as being chiefly a matter of epistemic fruitfulness, generalisation power, etc (which + scientific realist commitments leads to seeing them as real too). Constraints I think definitely count as candidate posits with the required epistemic virtues in this regard.