I don’t know the concrete details about what “taking on a global value” looks like, but I visualize a grid (like in Kevin Simler’s going critical post) that has a few competing colors trying to spread, and it seems reasonable that you could tweak the setting of the network such that very quickly one signal dominates the entire network.
But I don’t know how to actually make something like that.
If you’re interested in the TIN specifically, what I got from the paper was “here’s a totally plausible candidate, and from what we know about self-organization in neural networks, it could totally do this functionality”.
The biggest reason to think that there’s something that’s winner-take-all with a global value, is to explain bottlenecks that won’t go away. Intentional conscious thought seems to be very serial, and the neural turing machine model does a decent jump of showing how a global workspace is central to this. If there’s no global workspace, and there’s just the thalamus doing sensory gating, and routing chunks of cortex to each other, I’d expect to see a lot more multi tasking ability.
Also, this is more a property than a constraint, if global communication works by routing then everything that’s routed needs to know where it’s going. This makes sense for some systems, but I think part of the cool flexibility in a GNW architecture is that all of the cortex sees the contents of the GNW, and subsystems that compute with that as an input can spontaneously arise.
If there’s no global workspace, and there’s just the thalamus doing sensory gating, and routing chunks of cortex to each other, I’d expect to see a lot more multi tasking ability.
What if there is global workspace, but it doesn’t hold one value? On some level it has to be true anyway—perception is not one-dimensional. And it all depends on definition (granularity) of task—if we need to explain why global workspace can’t be dominated by page with half math problems and half story, then we can use the same explanation for why the state of workspace learned to not usually be like that. I can see how interconnectedness of workspace means all parts of input vector influence all of the workspace’s state, and so you can’t easily process different inputs independently, but can’t you process combined input? Isn’t it what happens, when you first just see something, then hear “Tell me what you see”, and the action is produced because of what you see and hear?
Some sort of “combination” seems plausible for perception. Baars actually mentions “The binding problem” (how is it that disparate features combine to make a cohesive singular perception) but I couldn’t see how their idea addressed it.
This is actually one of the reasons I’m interested in looking for stuff that might be the “clock time” of any sort of bottleneck. Some amount of simultaneity of perception seems to be a post production thing. The psychological refractory period relates to experiments where you see and hear something and have to respond, and one seems to block the other for a moment (I haven’t investigated this in depth, so I’m not v familiar with the experimental paradigm). But there are other things that totally seem like simultaneously experience modalities of perception. I wonder what sorts of experiments would piece apart “actually happening at the same time” from “rapid concurrent switching + post production experience construction”. I’m very interested in finding out.
I don’t know the concrete details about what “taking on a global value” looks like, but I visualize a grid (like in Kevin Simler’s going critical post) that has a few competing colors trying to spread, and it seems reasonable that you could tweak the setting of the network such that very quickly one signal dominates the entire network.
But I don’t know how to actually make something like that.
If you’re interested in the TIN specifically, what I got from the paper was “here’s a totally plausible candidate, and from what we know about self-organization in neural networks, it could totally do this functionality”.
The biggest reason to think that there’s something that’s winner-take-all with a global value, is to explain bottlenecks that won’t go away. Intentional conscious thought seems to be very serial, and the neural turing machine model does a decent jump of showing how a global workspace is central to this. If there’s no global workspace, and there’s just the thalamus doing sensory gating, and routing chunks of cortex to each other, I’d expect to see a lot more multi tasking ability.
Also, this is more a property than a constraint, if global communication works by routing then everything that’s routed needs to know where it’s going. This makes sense for some systems, but I think part of the cool flexibility in a GNW architecture is that all of the cortex sees the contents of the GNW, and subsystems that compute with that as an input can spontaneously arise.
What if there is global workspace, but it doesn’t hold one value? On some level it has to be true anyway—perception is not one-dimensional. And it all depends on definition (granularity) of task—if we need to explain why global workspace can’t be dominated by page with half math problems and half story, then we can use the same explanation for why the state of workspace learned to not usually be like that. I can see how interconnectedness of workspace means all parts of input vector influence all of the workspace’s state, and so you can’t easily process different inputs independently, but can’t you process combined input? Isn’t it what happens, when you first just see something, then hear “Tell me what you see”, and the action is produced because of what you see and hear?
Some sort of “combination” seems plausible for perception. Baars actually mentions “The binding problem” (how is it that disparate features combine to make a cohesive singular perception) but I couldn’t see how their idea addressed it.
This is actually one of the reasons I’m interested in looking for stuff that might be the “clock time” of any sort of bottleneck. Some amount of simultaneity of perception seems to be a post production thing. The psychological refractory period relates to experiments where you see and hear something and have to respond, and one seems to block the other for a moment (I haven’t investigated this in depth, so I’m not v familiar with the experimental paradigm). But there are other things that totally seem like simultaneously experience modalities of perception. I wonder what sorts of experiments would piece apart “actually happening at the same time” from “rapid concurrent switching + post production experience construction”. I’m very interested in finding out.