250 likes | 262 Vues
Explore the integration and advancement of anticipatory cognitive mechanisms in building Cognitive Systems with predictive abilities and emotional reactions to future scenarios.
E N D
Mind RACESFirst Review MeetingIntroduction Rino Falcone - Project Coordinator Institute of Cognitive Sciences and Technologies National Research Council of Italy MindRACES, First Review Meeting, Lund, 11/01/2006
Integration and Interaction Advancement Integration and Interaction Advancement Advancement Meeting Schedule • Introduction • Activities on the Scenarios (WP2) • Activities of advancement on anticipation • Activities on the thematic workpackages (WP3, WP4, WP5) • Activities about the Integration (WP6) • Robotic Demo • Simulation Demo • Some challenging questions Comparison The frontiers MindRACES, First Review Meeting, Lund, 11/01/2006
Cognitive Systems and Anticipation From the Objective of the 'Cognitive Systems’ Action Line: “To construct physically instantiated or embodied systems that can perceive, understand and interact with their environment, and evolve in order to achieve human-like performance in activities requiring context-(situation and task) specific knowledge” It is really difficult to think about systems performing like (and interacting with) humans without any mechanism of anticipation Only a cognitive system with anticipation mechanisms can be credible, adaptive, and successful in interaction with the environment and in social interaction with other autonomous systems and with humans MindRACES, First Review Meeting, Lund, 11/01/2006
Walking with your glass of wine Dark room Falling down breaks To represent and to work on the representation Recognize Rebuilt the possible mental perceptions of these pieces Take the pieces of glass The Anticipatory Nature of Mind MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES Goal The general goal of the Mind RACES project is to investigate different anticipatory cognitive mechanisms and architectures, their functionalities and integrations in order to build Cognitive Systems endowed with the ability: - to predict the outcome of their actions, - to build a model of future events, - to control their perception anticipating future stimuli and - to emotionally react to possible future scenarios MindRACES, First Review Meeting, Lund, 11/01/2006
Attention Perception Anticipation Goal-oriented behavior Emotion Attention Perception Goal-oriented behavior Emotion Cognitive Systems and Anticipation In many other works (also in ), anticipation was implicitly and indirectly managed Anticipation MindRACES, First Review Meeting, Lund, 11/01/2006
Prediction and Anticipation An anticipatory system is a system containing a predictive model of itself and/or of its environment that allows it to change state at an instant in accord with the model’s predictions pertaining to a later instant (Robert Rosen) Prediction is a representation that a particular event will occur in the future Anticipation is a future-oriented action/decision/behavior based on a (implicit or explicit) prediction MindRACES, First Review Meeting, Lund, 11/01/2006
Purposes of Anticipation Anticipatory processes serve many purposes: - increasing agent’s promptness to react - raising the possibility of success in managing the huge amount of stimuli and information we receive - aiding autonomous agents in interaction and coordination with the environment and with other agents MindRACES, First Review Meeting, Lund, 11/01/2006
Prediction is necessary for evaluating future world states/situations Anticipatory actions/decisions/behaviours for obtaining goals, performing intended/adapted behaviors and so on Prediction Anticipation The agent could (has to) make new predictions (less approximates than before): adjustments, less approximation Cycle Prediction-Anticipation MindRACES, First Review Meeting, Lund, 11/01/2006
The Prediction bases • Low level 'predictions' are based on some form of 'statistical' learning (on frequency and regular sequences, on judgment of normality in direct perceptual experience), on the strength of associative links and on the probability of activation (Kahneman & Miller 1986; Ortony & Partridge 1987). • High level predictions have many different sources: • Analogy (“The first time he was very elegant, I think that he will be well dressed”) and, in general, • Inferences and Reasoning (“He is Italian thus he will love pasta”), to natural laws and frame problem; and - in social domain - • Norms, Roles, Conventions, Habits, Scripts (“He will not do so; here it is prohibited”), • Theory of Mind (“He hate John, so he will try to…”; “He decided to go in vacation, so he will not be here on Monday”). • Emotion-based predictions • Emotions eliciting an anticipatory behavior(for examplearousal) • Emotions eliciting predictions (for example fear eliciting belief of danger) MindRACES, First Review Meeting, Lund, 11/01/2006
! Event approaching predator Kinds of Anticipation Explicit Anticipation(the organism/system is able to generate “representations” of the forthcoming events at different time scales) Implicit Anticipationor behavioral anticipation(the response is associated with a stimulus (a precursor). The behavior is selected to react to the event that is forthcoming (preparatory behavior) Precursory stimulus e.g.: noise Preparatory behavior jump MindRACES, First Review Meeting, Lund, 11/01/2006
Psychology of Action Artificial Intelligence Situated and Developmental Robotics Mathematics, Adaptive Robotics Cognitive Modelling, Agent & MAS theory Evolutionary and Bio-Inspired Robotics, Artificial Life Artificial Intelligence Believable agents, Affective Computing Situated Developmental Robotics Cognitive Modelling Cognitive Modelling Consortium Expertise on Anticipation Partner Expertise on Anticipation Anticipatory Behavioral Control UW-COGSCI OFAI Interactive Categorization, Symbol Grounding, Learning Optimal Prediction, Learning, Attention IDSIA Expectation theory, Trust theory, Emotions Neural Bases of Action and Planning ISTC Emotions, Affective Interaction IST Bio-inspired models of Attention, Learning LUND NBU Analogical Reasoning, Context NOZE Cognitive Architectures MindRACES, First Review Meeting, Lund, 11/01/2006
Project Characteristics - Inter-disciplinarity (Artificial Life, Artificial Intelligence, symbolic and sub-symbolic approach, Psychology, Biology, etc.) - Unification (coherent theory and ontology of anticipatory phenomena) - Multiple approachesto the Anticipation (from sensori-motor to social coordination) - FunctionalIntegration (from an architectural point of view to an effective integration of mechanisms) - Multiple approaches to the implementation and explorative experimentation of the models (conceptual, formal, simulative, robotic) MindRACES, First Review Meeting, Lund, 11/01/2006
Different Cognitive Functions MindRACES is studying anticipatory mechanisms in different Cognitive Functions: Attention, Monitoring, and Control (Work Package 3) Goal directed behaviour, Pro-activity, and Analogy (Work Package 4) Emotional Processes (Work Package 5) MindRACES, First Review Meeting, Lund, 11/01/2006
Cognitive Functions in WP3 Attention, Monitoring and Control (Work Package 3) The general goal of the cognitive system explored in this WP is to “understand” the world around it by focussing on the relevant details, hence the cognitive system is mainly motivated by an epistemic goal. The emphasis is on selective attention and knowledge representationviz. not on the use of knowledge for the guidance of action (to achieve pragmatic goals). Active perception: main focus in low-level perception. The cognitive system is able to perceive only by means of both sensors and actions that produce as effects new sensory signals (i.e. self-modification of the body orientation towards the stimuli). The cognitive system is attuned to changes in the sensory inputs. Constructive perception: is focussed on high-level perception where the sensorial stimulation is structured or organized through the application of schemas (recognition). The percepts are constructed through the schematization of the sensorial stimulation (sensations). The schemes are seen as procedures to construct and interpret the percepts. Anticipatory mechanisms:Expectation-based attention shifting, attention as epistemic control, constructive perception MindRACES, First Review Meeting, Lund, 11/01/2006
Cognitive Functions in WP4 Goal directed behaviour, Pro-activity and Analogy (Work Package 4) Cognitive system uses models of the world that take into consideration the consequences of its own action, that is models of the world that predict the future state of the world on the basis of current state and the planned action (also known as “forward models”). Anticipation for deciding what to do next: predictions are compared with a goal state not with a world/sensorial state; and are used for action/plan selection. Anticipation for action monitoring: expectations are compared with the actual state and the match/mismatch information is used e.g. for adjusting and tuning actions, generating surprise, shifting from a routine to a deliberative control, “compile” behaviours, etc. Anticipatory mechanisms: Sub-symbolic planning, pro-active activation of goals, anticipation at different time scales and levels of abstractions (i.e. anticipatory classifiers), construction of models of future events based on analogy MindRACES, First Review Meeting, Lund, 11/01/2006
Cognitive Functions in WP5 Anticipatory Emotions (Work Package 5) In this Project emotions are those bodily motions that are not simply information about bodily processes and states, but are ‘about’ an event or imagination, have a subjective qualitative value (are pleasant/unpleasant), and represent an implicit felt ‘evaluation’ of that event or idea (not based on reasons) Three different relationships between emotion and anticipation: i) Emotions eliciting an anticipatory (preparatory) behavior ii) Emotions due to Anticipatory Representations: a) Emotion now as a response to the predicted future event b) Emotion at the very predicted moment due to the previous expectation and its (mis)match with reality iii) Anticipating Future Emotions Anticipatory mechanisms:goals activation based anticipatory affective states (somatic markers), affective monitoring of goals’ satisfaction, appraisal of future events on the basis of perceived signs MindRACES, First Review Meeting, Lund, 11/01/2006
Scenarios’ Role • Function: The identification of the scenarios was based on the consideration that common scenarios should be considered as a mean: • to match different models, mechanisms and architectures • to achieve integration among the different approaches characterizing the partners • Objective: identify 6 scenarios corresponding to the situation in which cognition requires anticipation and implement 3 of them. • Three domains: • Anticipation in perception, selective attention and knowledge representation • Anticipation of the consequences of one’s own action for decisions and action monitoring • Anticipation and emotions MindRACES, First Review Meeting, Lund, 11/01/2006
Results after 1 year • - MindRACES assessed weaknesses and strengths of the main existing architectures in generating predictions and using those predictions for the generation of anticipatory cognitive functions (D4, D5, D6) • - MindRACES analyzed 27 scenarios. It identified and selected 6 of them corresponding to the situation in which cognition requires anticipation and we are going to implement 3 of them (D3) • - MindRACES realized advancements in software architecturesand robotic hardware useful for introducing mechanisms of anticipation (D7) • - MindRACES realized advancements in existing cognitive systems introducing mechanisms of prediction/anticipation (D7) • - Towards the integration: significance advances in sharing concepts, problems, methodologies among the various MindRACES approaches (D8) MindRACES, First Review Meeting, Lund, 11/01/2006
Practical Goals of MindRACES 1) incorporate anticipatory functionalities into existing cognitive models; 2) improve anticipatory functionalities of existing cognitive models 3) integrate different anticipatory functionalities of cognitive models MindRACES, First Review Meeting, Lund, 11/01/2006
Action Controlbased on forward and inverse modeling Hierarchical Architecture Action Control; interactive categorization based on anticipation Specification of mechanisms for development inspired by biological systems Evolino: EVolution of recurrent systems with Optimal LINear Output General Purpose, in particular Selective Attention Hierarchical Schema Mechanism, command fusion Action Controlbased on forward and inverse modeling Building Macro Actions Achievements: improve anticipatory functionalities of existing cognitive models Partner Improvement Functionality Starting Model UW XACS2- Anticipatory Classifiers Artificial Immune System (AIS) OFAI Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM) IDSIA ISTC/Noze Schema Mechanism ISTC Neural Networks, reinforcement learning MindRACES, First Review Meeting, Lund, 11/01/2006
Anticipatory Attention Control Anticipatory Attention Control Analogy based Anticipation Analogy based Anticipation Predicting the emotive potential of salient stimuli Emotivector Introduction of the Expectation All the uses of expectation (ex. Surprise) The Schema(s) generating better predictions are selected for Action Control Hierarchical, fuzzy-based Schema Mechanism Achievements: Incorporate anticipatory functionalities into existing cognitive models Partner Improvement Functionality Starting Model LUCS Visual Attention in IKAROS NBU Associative Memory Based Reasoning (AMBR) IST Believable character architecture Belief Desire Intention Architecture (BDI) ISTC ISTC/Noze Agent-based cognitive architecture AKIRA MindRACES, First Review Meeting, Lund, 11/01/2006
First steps towards Integration - Theoretical analysis of the MindRACES models and architectures from the perspective of possible integration of anticipatory mechanisms. MindRaces has developed an integration analysis: it has been made a systematic analysis of the possible integration needed to overcome successfully difficult aspects of the scenarios - Selection of promising integration and - Implementationin integrated models The integration of mechanisms of any kind is considered worthwhile if it improves non-trivially the performance of the model compared to a model using a single mechanism or a formal combination of two mechanisms MindRACES, First Review Meeting, Lund, 11/01/2006
Dissemination - Confrontation In addition to the traditional events about the project dissemination (Conferences and workshops, Publications in scientific journals, International workshop about Anticipatory Systems, Interviews, Project Presentation, etc.) AAAI 2005 Fall Symposium From Reactive to Anticipatory Cognitive Embodied Systems November 3-6, 2005, Hyatt Crystal City in Arlington, Virginia Presentation of MindRACES Project (Mark Bickhard, Stephen Grossberg , Andrew Ortony, Deb Roy ) MindRACES, First Review Meeting, Lund, 11/01/2006
Scenarios High Cognition Analogy Emotions Learning Attention Demos Integration Challenges Blueprint of the day Introduction Themes and Approaches MindRACES, First Review Meeting, Lund, 11/01/2006