The Simple Dynamic Cognition Cycle flow diagram above is topologically identical to this one:
I published with an earlier post on the Salience theory of dynamic cognition and consciousness but this one shows the multiple feedback lines from the salience node as all left facing. The only one that is right facing feeds to Action.
This is important because it shows that Salience evaluations directly modulate Action while also bypassing action for continued Sensation, some times a prediction is incorrect and is bypassed for continued approximation of stored memory with new sensation in the Comparison node which takes feedback from salience itself (emotional and autonomic).
This is important because it shows that Salience evaluations directly modulate Action while also bypassing action for continued Sensation, some times a prediction is incorrect and is bypassed for continued approximation of stored memory with new sensation in the Comparison node which takes feedback from salience itself (emotional and autonomic).
I posit that cognition (the mind) happens between Sensation and Salience nodes and some times bubbles up Action, the "self" and consciousness are emergent reflections of this sea of comparisons in real time.
When we sleep the outer red flow arrows from salience to action are minimally triggered but subconscious cognition can still continue as memories are recalled for continued sub sensory evaluation, after all memories are basically copies of incoming sensation stored to make predictions as modulated by salience (tagged with emotional and autonomic import ratings) so if the Action is not driven that doesn't mean that comparison and salience evaluation aren't still on going particularly since autonomic salience must continue to be monitored as the agent "sleeps".
In the mammalian brain sleep seems to be important for low level memory consolidation and organization activities that would be inefficient to do during the wake state (they'd mar cognitive performance for obvious reasons you would be adding new sensations while trying to consolidate old ones). It should be possible to create an artificial cognition that doesn't need sleep (or needs less of it) by allowing consolidation to happen in parallel using an independent computational engine from the one processing real time sensation.
I hypothesize that in a dynamic cognition that correctly mirrors biological cognitive flow if an attempt to consolidate memories is not made the efficiency of cognitive processes will steadily degrade over time this would be due to the over loading of action delta data in the early virtual neuron layers of the sensory comparison and storage stacks for each dimensional modality.
When we sleep the outer red flow arrows from salience to action are minimally triggered but subconscious cognition can still continue as memories are recalled for continued sub sensory evaluation, after all memories are basically copies of incoming sensation stored to make predictions as modulated by salience (tagged with emotional and autonomic import ratings) so if the Action is not driven that doesn't mean that comparison and salience evaluation aren't still on going particularly since autonomic salience must continue to be monitored as the agent "sleeps".
In the mammalian brain sleep seems to be important for low level memory consolidation and organization activities that would be inefficient to do during the wake state (they'd mar cognitive performance for obvious reasons you would be adding new sensations while trying to consolidate old ones). It should be possible to create an artificial cognition that doesn't need sleep (or needs less of it) by allowing consolidation to happen in parallel using an independent computational engine from the one processing real time sensation.
I hypothesize that in a dynamic cognition that correctly mirrors biological cognitive flow if an attempt to consolidate memories is not made the efficiency of cognitive processes will steadily degrade over time this would be due to the over loading of action delta data in the early virtual neuron layers of the sensory comparison and storage stacks for each dimensional modality.
In a sequential neural network the rate at which new data is fed into the system for training is deterministic and fixed, once trained new evaluations happen independent of continuous training, but in a real cognitive agent one has no control over when sensory data is coming into the system and one must handle it in that moment...doing so re weights low level network layers and thus reducing the efficiency of all the cognitive dimensional comparison tasks associated with those layers. The ability to accurately make predictions goes down as this noise piles up in lower layers of the cognitive stack.
So I posit that some kind of offline consolidation will be needed to push sensed memories deeper into the virtual neuron stack and thus allowing future predictions to be more accurate, assuming of course that the DCC above is the only valid control flow that can emerge "mind". A different control flow and implementing architecture may be able to forgo this need.
The fact that the control flow described above between the 4 nodes may be a sign that it is correct as it seems to replicate the need for an analog of sleep. Again this is the simple diagram, the complex diagram has some particularly important symmetries identified that define how the cognitive engine would precisely flow between the different sensory modalities...getting those flows correct from the salience node is critical to emerging cognition. I will not be publishing the complex flow as I plan on working to implement it over the next 10 years or so...the significance of the simple diagram is that it presents the key innovation of salience evaluation as a fundamental requirement of dynamic cognition while describing exactly where it goes in the cognitive flow diagram.
Comments