Difference between revisions of "Cognitive Architecture Ideas"

From Wiki for iCub and Friends
Jump to: navigation, search
(Added comment from Claes von Hofsten about the distinction between locomotion, reaching, and grasping; about innate exogenous saliences; and about objecthood)
 
(5 intermediate revisions by 2 users not shown)
Line 1: Line 1:
* Affective state should influence more than just action selection, e.g.
+
* Affective state should influence more than just action selection and should be an aspect of the associative memory so that emotions are associated with events.
** an aspect of the associative memory (so that emotions are associated with events)
+
** ''Comment:'' this feature will be added in the future.
** a modulator of salience (so that certain features can be modulated by motivations)
+
 
 +
 
 +
* Affective state should also be a modulator of salience so that certain features can be modulated by motivations.
 +
** ''Comment:'' It is not clear how this can be effected directly.  Instead, once affective state is incorporated into the auto-associated episodic memory, and by extension in the procedural memory, affective state will implicitly modulate salience.
 +
 
 +
 
 
*Clarity is needed on distinction between auto-associative and hetero-associative memory modules. A better distinction is between episodic memories, procedural memories, and generalized world modeling. All of these should probably be multimodal with unimodal subnetworks.  
 
*Clarity is needed on distinction between auto-associative and hetero-associative memory modules. A better distinction is between episodic memories, procedural memories, and generalized world modeling. All of these should probably be multimodal with unimodal subnetworks.  
 +
** ''Comment:'' This suggestion has been taken up throughout the current version of the [[iCub Cognitive Architecture]].
 +
 +
 
*It is not clear to me that there is a fundamental distinction between locomotion, reaching, and grasping. Although the inputs are different (e.g. vestibular system in locomotion and haptic feedback in grasping), the fundamental algorithm might be the same in the brain. All relate to the goal of changing the robots position to a desired location in space. For example, brachiation requires all three of these tasks to be performed as a single coordinated motion.
 
*It is not clear to me that there is a fundamental distinction between locomotion, reaching, and grasping. Although the inputs are different (e.g. vestibular system in locomotion and haptic feedback in grasping), the fundamental algorithm might be the same in the brain. All relate to the goal of changing the robots position to a desired location in space. For example, brachiation requires all three of these tasks to be performed as a single coordinated motion.
 +
** ''Comment:'' The distinction between them is purely functional. Locomotion, reaching, and grasping all utilize inputs from all senses [Claes von Hofsten].
 +
 +
 
*Action selection should maybe extend to things other than the hetero-associative memory, locomotion, and reach/grasp. Like affective state, it should probably be able to send signals to much of the system.
 
*Action selection should maybe extend to things other than the hetero-associative memory, locomotion, and reach/grasp. Like affective state, it should probably be able to send signals to much of the system.
 +
** ''Comment:'' this feature will be added in the future.
 +
 +
 
*Robot is "born" with only exogenous salience, e.g. closeness, loudness, bright colors, etc. However, as a model is developed of the world endogenous salience comes from looking for known objects in the environment.
 
*Robot is "born" with only exogenous salience, e.g. closeness, loudness, bright colors, etc. However, as a model is developed of the world endogenous salience comes from looking for known objects in the environment.
 +
** ''Comment:'' The robot should be born with some other saliences like biological motion and face ... [Claes von Hofsten].
 +
 +
 
*Speaking of objects, we talked about where the notion of objects comes into the cognitive architecture. Rightly so, there is no inherent concept of objecthood. However, objects emerge from closely associated inputs (objects in the loose sense) or learning about the coherence of physical things (objects in a narrower sense).
 
*Speaking of objects, we talked about where the notion of objects comes into the cognitive architecture. Rightly so, there is no inherent concept of objecthood. However, objects emerge from closely associated inputs (objects in the loose sense) or learning about the coherence of physical things (objects in a narrower sense).
 +
** ''Comment:'' There is an inherent concept of objecthood in most animals. That is a notion of entities that have inner unity and outer borders that move  on continuous trajectories and exist over time [Claes von Hofsten].
 +
 +
 +
*Procedural memory needs to be a generative model which can "imagine" consequences of actions and predict future observations. This is likely very important for object affordances that we have discussed.
 +
** ''Comment:'' This capability is anticipated in the  [[Hetero-associative Procedural Memory Specification]] and will be added to the procedural memory in the future.

Latest revision as of 10:41, 15 August 2009

  • Affective state should influence more than just action selection and should be an aspect of the associative memory so that emotions are associated with events.
    • Comment: this feature will be added in the future.


  • Affective state should also be a modulator of salience so that certain features can be modulated by motivations.
    • Comment: It is not clear how this can be effected directly. Instead, once affective state is incorporated into the auto-associated episodic memory, and by extension in the procedural memory, affective state will implicitly modulate salience.


  • Clarity is needed on distinction between auto-associative and hetero-associative memory modules. A better distinction is between episodic memories, procedural memories, and generalized world modeling. All of these should probably be multimodal with unimodal subnetworks.


  • It is not clear to me that there is a fundamental distinction between locomotion, reaching, and grasping. Although the inputs are different (e.g. vestibular system in locomotion and haptic feedback in grasping), the fundamental algorithm might be the same in the brain. All relate to the goal of changing the robots position to a desired location in space. For example, brachiation requires all three of these tasks to be performed as a single coordinated motion.
    • Comment: The distinction between them is purely functional. Locomotion, reaching, and grasping all utilize inputs from all senses [Claes von Hofsten].


  • Action selection should maybe extend to things other than the hetero-associative memory, locomotion, and reach/grasp. Like affective state, it should probably be able to send signals to much of the system.
    • Comment: this feature will be added in the future.


  • Robot is "born" with only exogenous salience, e.g. closeness, loudness, bright colors, etc. However, as a model is developed of the world endogenous salience comes from looking for known objects in the environment.
    • Comment: The robot should be born with some other saliences like biological motion and face ... [Claes von Hofsten].


  • Speaking of objects, we talked about where the notion of objects comes into the cognitive architecture. Rightly so, there is no inherent concept of objecthood. However, objects emerge from closely associated inputs (objects in the loose sense) or learning about the coherence of physical things (objects in a narrower sense).
    • Comment: There is an inherent concept of objecthood in most animals. That is a notion of entities that have inner unity and outer borders that move on continuous trajectories and exist over time [Claes von Hofsten].


  • Procedural memory needs to be a generative model which can "imagine" consequences of actions and predict future observations. This is likely very important for object affordances that we have discussed.