Artificial Cognitive Systems 3 1 Introduction to Cognitive Robotics Introduction to Cognitive Robotics Module 6: Artificial Cognitive Systems Lecture 3: Memory and prospection; internal simulation, the symbol grounding problem www.cognitiverobotics.net
59
Embed
Introduction to Cognitive RoboticsTypes of Memory Episodic •Episodic memory is a constructive process –Each time an event is assimilated into episodic memory, past episodes are
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Artificial Cognitive Systems 3 1 Introduction to Cognitive Robotics
Introduction to Cognitive Robotics
Module 6: Artificial Cognitive SystemsLecture 3: Memory and prospection; internal simulation, the symbol grounding problem
www.cognitiverobotics.net
Artificial Cognitive Systems 3 2 Introduction to Cognitive Robotics
Memory
• Memory plays a crucial and sometimes unexpected role in cognition
• Strong parallel between memory and knowledge
• Memory and knowledge are equivalent: they both encapsulate the experience that arises from interaction with the world
• Memory is not just as a passive mechanism for storing knowledge about the past
Artificial Cognitive Systems 3 3 Introduction to Cognitive Robotics
Types of Memory
• Declarative
• Procedural
• Semantic
• Episodic
• Long-term
• Short-term
• Working
• Modal
• Amodal
• Symbolic
• Sub-symbolic
• Hetero-associative
• Auto-associative
Artificial Cognitive Systems 3 4 Introduction to Cognitive Robotics
Types of Memory
Declarative memory
• Knowledge of things / facts
• “Knowing that” • Propositional memory (true or false)
• Can be communicated from one agent to another through language
• Can be acquired in a single act of perception or cognition
• Accessible to conscious recall
• Explicit memory
Artificial Cognitive Systems 3 5 Introduction to Cognitive Robotics
Types of Memory
Procedural memory
• Skill-oriented memory of actions
• “Knowing how”• Can only be demonstrated
• Acquired progressively and may require an element of practice
• Not accessible to conscious recall
• Implicit memory
• Non-declarative memory
Artificial Cognitive Systems 3 6 Introduction to Cognitive Robotics
Types of Memory
Episodic memory
• Specific instances in the agent’s experience: autobiographical
• Explicit spatial and temporal context– what happened, where it happened, and when it happened
– This temporal sequencing is the only element of structure in episodic memory
• Sub-symbolic
Artificial Cognitive Systems 3 7 Introduction to Cognitive Robotics
Types of Memory
Episodic memory
• Episodic memory is a constructive process
– Each time an event is assimilated into episodic memory, past episodes are re- constructed a little differently each time
– Related to the role that episodic memory plays in the process of internal simulation that forms the basis of prospection
Artificial Cognitive Systems 3 8 Introduction to Cognitive Robotics
Types of Memory
Semantic memory
• General knowledge about the agent’s world: facts, ideas, and concepts
• May be independent of the agent’s specific experiences
• Memory necessary for the use of language
• Derived from episodic memory through a process of generalization and consolidation • Symbolic
Artificial Cognitive Systems 3 9 Introduction to Cognitive Robotics
Types of Memory
Modal memory
• Tied directly to a particular sensory modality such as vision, audition, or touch
• Episodic memory though is more likely to be modal since it is closely tied to an agents’s specific experiences
Amodal memory
• Amodal memory has no necessary association with the sensorimotor experiences
• Semantic declarative facts, represented symbolically, are typically amodal
Artificial Cognitive Systems 3 10 Introduction to Cognitive Robotics
Types of Memory
Associative memory
• An element of information or some pattern is linked to another
• The first element or pattern is used to recall the second, by association
AssociativeMemory
Pattern A Pattern B
Artificial Cognitive Systems 3 11 Introduction to Cognitive Robotics
Types of Memory
Hetero-associative memory
• Recalls a memory that is different in character from the input
• A particular smell or sound, for example, might evoke a visual memory of some past event
AssociativeMemory
Sound Image
Artificial Cognitive Systems 3 12 Introduction to Cognitive Robotics
Types of Memory
Auto-associative memory
• Recalls a memory of the same modality as the one that evoked it
• A picture of a favourite object might evoke a mental image of that object in vivid detail
AssociativeMemory
Image A Image B
Artificial Cognitive Systems 3 13 Introduction to Cognitive Robotics
Role of Memory
The role of memory – why do we remember things?
– To recognize objects, events, and people we’ve encountered before
– To act towards them in some appropriate way (attraction/avoidance)
– Memory is what makes it possible for the changes that occur as a result of learning and development to persist
– Memory also is what makes it possibleto project forwards into the future
Artificial Cognitive Systems 3 14 Introduction to Cognitive Robotics
Role of Memory
One of the central pillars of cognitive capacity:
– the ability to simulate internally the outcomes of possible actions and select the most appropriate one for the current situation
– Memory can be seen as a mechanism that allows a cognitive agent to prepare to act,overcoming through anticipation the inherent “here-and-now” limitations of its perceptual capabilities
– a cognitive system doesn’t operate just on the basis of its current sensory data but readies itself for what it expects and adjusts to the unexpected
Artificial Cognitive Systems 3 15 Introduction to Cognitive Robotics
Role of Memory
Memory is an active & constructive process, and it is fundamentally associative
– Memories are recalled by associated triggers, possibly other memories
– If you have a network of associative memories, you can run through this network backwards or forwards
– Running through it forwards provides the anticipatory predictive element of memory suggesting possible sequence of events leading to a desired goal
– Running through it backwards provides a way of explaining how some event or other might have occurred or imagining ways in which it might have turned out differently
Artificial Cognitive Systems 3 16 Introduction to Cognitive Robotics
“It’s a poor sort of memory that only works backwards”
Remarks of the White Queen to Alicein Lewis Carroll’s Through the Looking Glass
Memory is Prospective
Artificial Cognitive Systems 3 17 Introduction to Cognitive Robotics
“It’s a poor sort of memory that only remembers what has actually happened”
Remarks by Tom Ziemke
at a talk in Linköping University
Memory is Constructive
Artificial Cognitive Systems 3 18 Introduction to Cognitive Robotics
– Old episodic memories are reconstructed slightly differently every time a new episodic memory is assimilated or remembered
– The constructive episodic simulation hypothesis [Schacter and Addis 2007]
• Episodic memory allows the simulation of multiple possible futures
• This imposes an even greater need for a constructive capacity because of the need to extrapolate beyond past experiences
Artificial Cognitive Systems 3 25 Introduction to Cognitive Robotics
Artificial Cognitive Systems 3 26 Introduction to Cognitive Robotics
The Past
Past events arereconstructed ...
Episodic MemorySpecific instances of
the agents experience
Artificial Cognitive Systems 3 27 Introduction to Cognitive Robotics
The Past
Past events arereconstructed ...
Episodic Memory
To allow the agent to pre-experience the future
The Future
Artificial Cognitive Systems 3 28 Introduction to Cognitive Robotics
The Past
Past events arereconstructed ...
Episodic Future Thinking
To allow the agent to pre-experience the future
The Future
C. M. Atance and D. K. O’Neill, “Episodic future thinking,” Trends in Cognitive Sciences, vol. 5, no. 12, pp. 533–539, 2001.
Artificial Cognitive Systems 3 29 Introduction to Cognitive Robotics
The Past
Past events arereconstructed ...
The constructive episodic simulation hypothesis
To allow the agent to pre-experience the future
Future B
D. L. Schacter and D. R. Addis, “The cognitive neuroscience of constructive memory: Remembering the past and imagining the future,” Philosophical Transactions of the Royal Society B, vol. 362, pp. 773–786, 2007.
Future A
Future C...
Artificial Cognitive Systems 3 30 Introduction to Cognitive Robotics
The Future
SemanticEpisodic
K. K. Szpunar, R. N. Spreng, and D. L. Schacter, A taxonomy of prospection: introducing anorganizational framework for future-oriented cognition, PNAS 111(52), 18414–18421, 2014.
General knowledge about the agent's world
Artificial Cognitive Systems 3 31 Introduction to Cognitive Robotics
The Future
SemanticEpisodic
Procedural
K. K. Szpunar, R. N. Spreng, and D. L. Schacter, A taxonomy of prospection: introducing anorganizational framework for future-oriented cognition, PNAS 111(52), 18414–18421, 2014.
Action knowledge and motor skills
Artificial Cognitive Systems 3 32 Introduction to Cognitive Robotics
This elicits a covert action which in turn elicits a simulated perception and a consequent covert action
Artificial Cognitive Systems 3 44 Introduction to Cognitive Robotics
Internal Simulation and Action
HAMMER accomplishes internal simulation using forward and inverse models
– The inverse model
• Takes as input the current state of the system and the desired goal, and it outputs the motor commands necessary to achieve that goal
– The forward model
• Acts as a predictor
• Takes as input the motor commands and simulates the perception that would arise if this motor command were to be executed
(just as the simulation hypothesis envisages)
ActionPercept
Goal
PerceptAction
Artificial Cognitive Systems 3 45 Introduction to Cognitive Robotics
Y. Demiris and B. Khadhouri. Hierarchical attentive multiple models for execution and recognition (HAMMER). Robotics and Autonomous Systems, 54:361– 369, 2006.
Inverse Model Forward ModelInverse Model Forward Model
Perception Action Perception
Internal Simulation and Action
HAMMERArchitecture
Artificial Cognitive Systems 3 46 Introduction to Cognitive Robotics
Inverse Model Forward ModelInverse Model Forward Model
Perception Action Perception
Internal Simulation and Action
HAMMERArchitecture
Artificial Cognitive Systems 3 47 Introduction to Cognitive Robotics
The HAMMER Architecture[Demiris and Khadhouri 2006]
Y. Demiris and B. Khadhouri. Hierarchical attentive multiple models for execution and recognition (HAMMER). Robotics and Autonomous Systems, 54:361– 369, 2006.
Internal Simulation and Action
Artificial Cognitive Systems 3 48 Introduction to Cognitive Robotics
The HAMMER Architecture[Demiris and Khadhouri 2006]
The choice of inverse/forward model pair is made by an internal attention process based on how close the predicted outcome is to the desired one
Y. Demiris and B. Khadhouri. Hierarchical attentive multiple models for execution and recognition (HAMMER). Robotics and Autonomous Systems, 54:361– 369, 2006.
Internal Simulation and Action
Artificial Cognitive Systems 3 49 Introduction to Cognitive Robotics
D. Wolpert, R. C. Miall, and M. Kawato. Internal models in the cerebellum. Trends in Cognitive Sciences, 2(9):338–347, 1998.
Cognitive Behavior of Humans, Animals, and Machines: Situation Model Perspectives 49 Center for Interdisciplinary Research, University of Bielefeld, Germany.
Multiple pairs of inverse & forward models
Internal Simulation and Action
Artificial Cognitive Systems 3 50 Introduction to Cognitive Robotics
Internal Simulation and Action
• HAMMER provides for the hierarchical composition of primitive actions into more complex sequences
• It has been implemented both in robot simulations and on physical robotic platforms
Artificial Cognitive Systems 3 51 Introduction to Cognitive Robotics
Internal Simulation and Action
• HAMMER goes beyond the scope of episodic memory in effecting internal simulation by invoking actions and behaviours
• The sensorimotor associations involved in internal simulations, for forward and inverse models, requires both episodic memory and procedural memory
• Episodic memory is needed for visual imagery, including proprioceptive imagery
• Procedural memory is needed for motor imagery
Artificial Cognitive Systems 3 52 Introduction to Cognitive Robotics
The Symbol Grounding Problem
• Assuming a cognitive system has some form of symbolic representation of the world around it
– i.e. some set of tokens that denote objects in the agent’s world
• How does the representation, i.e. the symbols, acquire meaning?
How do purely symbolic representations acquire semantic content?
• These apparently innocent questions are made difficult by the fact that (physical) symbol systems are governed by purely syntactic processes
Artificial Cognitive Systems 3 53 Introduction to Cognitive Robotics
The Symbol Grounding Problem
• Physical symbol systems
– Atomic symbols
– Strings of symbols
– Symbol-based rules that define the manipulation and recombination of symbols and strings of symbols
• Defined in terms that make no reference to what these symbols mean
Artificial Cognitive Systems 3 54 Introduction to Cognitive Robotics
The Symbol Grounding Problem
• But they are all “semantically interpretable”
– the syntax can be assigned a semantic meaning so that symbols and strings of symbols can represent objects, events, or concepts, and describe them or stand in for them
• The problem is how to assign this meaning
• This is the symbol grounding problem [Harnad 1990]
Artificial Cognitive Systems 3 55 Introduction to Cognitive Robotics
The Symbol Grounding Problem
• Symbolic representations have to be grounded bottom-up in non-symbolic representations of two kinds:
• Iconic representations– Derived directly from sensory data (e.g. visual imagery, motor imagery)
– Allow you to discriminate between different objects
• Categorical representations– Based on the output of both learned and innate processes that detect invariant features of
object & event categories from these sensory data(e.g. object affordances, repeated behaviours ... food is nice; dogs bite)
Artificial Cognitive Systems 3 56 Introduction to Cognitive Robotics
The Symbol Grounding Problem
• Higher-order symbolic representations can then be derived from these elementary symbols
• Both types of representation are non-symbolic
• A non-symbolic process is required to learn the invariances and thereby form the categories
Artificial Cognitive Systems 3 57 Introduction to Cognitive Robotics
The Symbol Grounding Problem
• Usually, we use some form of connectionist approach
– Create the mapping
– Form the categorical representation
• As a consequence, according to this argument, a grounded symbol system is a hybrid system: a combination of symbolic and emergent approaches (h not H)
Artificial Cognitive Systems 3 58 Introduction to Cognitive Robotics
The Symbol Grounding Problem
• A restricted form is the symbol anchoring problem[Coradeschi and Saffiotti 2003]
• Concerned only with artificial systems
Artificial Cognitive Systems 3 59 Introduction to Cognitive Robotics
The Symbol Grounding Problem
• Establish a relationship between a symbolic label denoting some object and the sensory perception of that object
• Maintaining that relationship over extended periods of time, even when that object cannot be seen
• Only concerned only with grounding physical objects
– Doesn’t address the grounding of abstract concepts (war or peace)