1 three partitioned-model_unifi_cnr

234 views
197 views

Published on

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
234
On SlideShare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
2
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide
  • \n
  • \n
  • \n
  • \n
  • \n
  • \n
  • \n
  • \n
  • \n
  • The input pattern weight the external information (Activation Score) and its relevance is given by the factorial score obtained weighting the internal knowledge (context). If activated the A-Scheme modifies the Knowledge (K) with the Extracted Factors (F) \nThe goal scheme (GS) is activated according to its Factorial Activation Score (based on K). GS modifies K which can cause it to deactivate. The Goal “Emotional” Factors are used to choose the appropriate B-Scheme.\nThe B-Scheme is activated depending both to its Factorial Activation Score and to the overlapping between the B-Extracted Factors and the Goal Emotional Factors. Also the cost of the scheme is considered as scheme selecting criterion. \n
  • \n
  • \n
  • Esempio VALIDO SOLO in laboratorio, SCHEMA come PERCETTRONE nella relatà i processi preattentivi distortcono grandemente tutti i tipi di percezione/detezione dell’informazione. In altri termini normalmente scatta la relevance heuristic ... questo è un esempio per far vedere come inizia il processo.\n
  • \n
  • For instance, if the context is "being at lunch", one may expect to see a given set of tools (fork, knife, etc.), even if they have an unusual shape, and not to see other things. If the scheme confirms this, the context is reinforced and one can activate schemes like "take the fork". If expected objects are missing or unexpected objects are present, the context is more dubious and other inputs schemes, like "process what is in the background" or "consider what you are earing" are activated. \n
  • \n
  • \n
  • \n
  • \n
  • ... si potrebbe far vedere come spesso i goal della gente siano fantasiosi ed irrealizzabili, pur rimanendo in grado di influenzare il loro comportamento. Quindi i GOAL non sottostanno necessariamente all’esame di realtà perchè possono essere approssimativi ed incompleti. Essi si limitano ad essere solo una “lista più o meno lunga di vincoli”\n
  • \n
  • \n
  • \n
  • \n
  • \n
  • \n
  • \n
  • 1 three partitioned-model_unifi_cnr

    1. 1. A tri-partite model of computational knowledge Andrea Guazzini° * University of Florence ° IIT - National Research Council of Italy (CNR) Funded by the EC FP7 Future Emerging Technologies Programme (Awareness), grant 257756 AWASS 2012 Edinburgh 10th-16th June 1
    2. 2. From humans to computerHumans have developed (through natural selection) “fast and frugal”methods for understanding the context, taking decisions and solving socialproblems in limited time and using bounded cognitive resources.These methods can have fruitful applications in ubiquitous computerappliances.Moreover, electronic devices are asked to interact with humans... and to act in their delegation. AWASS 2012 Edinburgh 10th-16th June 2 3
    3. 3. Self-awarenessHuman information processing is context-based.Human-computer interfaces are expected to behave in a personalized way,possibly extracting “sideways” information from geographic location, userprofiles, past interactions.But more information could be gathered by psychological analysis andcharacterization.Human-based heuristics can also result in more effective and optimizedsolutions for the typical case. AWASS 2012 Edinburgh 10th-16th June 3
    4. 4. Cognitive sciences fields Neural-level (neural networks) Functional areas and connections Experimental framework (factorial analysis) Dynamic behavior AWASS 2012 Edinburgh 10th-16th June 4 3
    5. 5. Digging into cognitionThere is a lot of knowledge accumulated by cognitive sciences, psychologyand sociology, but little is modeled in quantitative (and procedural) form.Most of modeling concerns basic functionalities (like the perceptivesystem), for instance using the ACTr scheme.There is a general agreement on different levels of information processing,related with response time and possibly with evolutive brain structures asrevealed by fMRI. We aim at implementing algorithmically the levels “below” rationalreasoning. AWASS 2012 Edinburgh 10th-16th June 5 3
    6. 6. The three levels of our modelingLet us suppose that the context is known. The modeling is performed accordingto this scheme: Perception encoding. Most of information in input behaves as noise (it is uncorrelated with the task – given the context). Reduction of information by projection on a subspace with limited number of dimensions. Representation and activation of knowledge. Implementation of action and redefinition of the context. Evaluation processes AWASS 2012 Edinburgh 10th-16th June 6 3
    7. 7. The minimal structure of a Self Awareness cognitive agentSelf awareness could be considered as an epiphenomenon of the cognitive processes of informationanalysis. Such processes can be classified on the basis of three criteria: Timescales, Cognitive Costsand Evolutionary features. Timescales -(Reaction times) Unconscious Knowledge (Perception and Pre-attentive activations)-> Fast (<.500 ms) Conscious Knowledge (Reasoning) -> Medium (From seconds to hours) Learning/Development -> Slow (From minutes to month) Cost (Cognitive Economy Principle - Amount of neural activation) Unconscious Knowledge -> Light (small and local activations) Conscious Knowledge -> Heavy (large and diffused activations) Learning/Development -> Very Heavy (diffused activations) Evolutionary features (Cognitive development) Unconscious Knowledge -> Critical period and “classical-hebbian” learning only (ACTr) Conscious Knowledge -> Trial and Error, Observation/Imitation and Induction learnings. Learning/Development -> Fixed hard wired rules. AWASS 2012 Edinburgh 10th-16th June 7
    8. 8. The minimal structure of a Self Awareness cognitive agent We denotes as schemes the procedure that manage information and perform actions, and by heuristics the management of scheme (activation, modification, learning). We divide schemes and heuristics in three modules: in the first one we put the structures that deal with input, in the second the actual management of information and actions and in the third the learning. This division is consistent with the the response time, but we think that there is a common structure of heuristics and schemes AWASS 2012 Edinburgh 10th-16th June 8 3
    9. 9. The minimal structure of a Self Awareness cognitive agentExternal Data Reaction time Module I Flexibility Unconscious knowledge perceptive and attentive processes Cognitive costs Relevance Heuristic Module II Reasoning Goal Heuristic Recognition Heuristic Solve Heuristic Module III Learning Behavior Evaluation Heuristic AWASS 2012 Edinburgh 10th-16th June 9
    10. 10. Some Unification Concepts A first step for a Mathematical translation Mental Schemes = knowledge Cognitive Heuristics = rules/functionsA-Scheme model: The input pattern weights the external information Extracted Activation (Activation Score) and its relevance is given by the Input pattern Factors (F) Factors (A) factorial score obtained weighting the internal knowledge (context). If activated the A-Scheme modifies the Knowledge (K) with the Extracted Factors (F)Goal-Schemes: The goal scheme (GS) is activated according to its Goal Factorial Activation Score (based on K). GSActivation Factors Extracted “Emotional” modifies K which can cause it to deactivate. The (G) Factors (F) Factors (E) Goal “Emotional” Factors are used to choose the appropriate B-Scheme.B-Scheme model: The B-Scheme is activated depending both to its FactorialActivation Factors Extracted Answer Activation Score and to the overlapping between the B- Factors (F) Extracted Factors and the Goal Emotional Factors. Also (B) & Cost the cost of the scheme is considered as scheme selecting criterion. Cognitive Heuristics Functions of distance estimation, correlation, minimization/maximization and combination among schemes. AWASS 2012 Edinburgh 10th-16th June 10
    11. 11. Module I: The unconscious knowledge From Gestalt to Relevance TheoryCognition is able to extract the relevant features from a given context“unconsciously”, integrating them continuously within the higher decisionalprocesses. e.g. the active process of perception (Data encoding) is the results of thecombination of the external information with the pre attentive activations.Involved cognitive processes Bottom Up processes which encode the information - e.g. Perception Top Down processes which filter the information - e.g. AttentionFundamental features Continuous detection and encoding of the incoming information Noise and dimensionality reduction of the information Updating of an associative representation of the context/environment (K) AWASS 2012 Edinburgh 10th-16th June 11
    12. 12. Module I: The unconscious knowledge From Gestalt to Relevance TheoryDynamics of MODULE I: Relevance Heuristic integrates the external information (EI) with the “pre attentive activations” (PA) in order to “choose” if activate a certain A-scheme. An A-scheme so can be characterized in terms of cognitive salience based on its overlapping with the vector (EI*PA) The activated A-Schemes are continuously accumulated in a multidimensional and sparse representation of the reality (Immanent Knowledge Vector - K). K integrates also projection from the module II. K is continuously analyzed by a factorial analysis, which drives the new steps of encoding/perception affecting the PA (weighting/selecting the new information - aka searching heuristics). Finally the Relevant Features (RF) for the next stages of the decisional process are extracted. AWASS 2012 Edinburgh 10th-16th June 12
    13. 13. Module I - Some Unification Concepts A-Scheme: The Knowledge Vector A-Schemes: knowledge building blocksInput Vector (I) I1 , I2 , ..., In Example: The KANITZA triangle (k) (k) (k) Scheme Sk W1 , W2 , ..., Wn Extracted Factors (Sk) Scheme activation score ⌦ (k) Module 1 deals with external information, which is multiform, huge and has to be filtered in order to focus on Example: The WORD recognition important components. The A-schemes do this, and extract information. They are "activated" by the score match of their input patterns with ROSE the context vector, they are validated by means of their relevance with the input, and, if accepted, they contribute to the context and pass information to schemes in module 2. A Flower The past of Rise AWASS 2012 Edinburgh 10th-16th June 13
    14. 14. Module I - Some Unification Concepts The Immanent Knowledge Vector, i.e. The context IKV: Immanent representation of the environment A-Schemes Example:T1 Silver Dish (S1) T2 Green Pocket (S2) T3 Fork & Glasses (S3,4) We assume that there is a structure that denotes the context frame, and we denote it as the Knowledge/ Context Vector. It is called vector since we assume that it represents the knowledge projected on a limited number of internal dimensions. The activated A-Schemes are continuously accumulated in a multidimensional and sparse representation of the reality (Immanent Knowledge - K). AWASS 2012 Edinburgh 10th-16th June 14
    15. 15. Module I - Some Unification Concepts The dimensionality reduction i.e. The pre-attentive processing RF: The relevant features used to activate the reference Context Frame A-Schemes Detected Context FrameT1 Silver Dish (S1) Example: A Set TableT2 Green Pocket (S2)T3 Fork & Glasses (S3,4) Schemes have an activation pattern, that can be modified at the learning level to "enhance" their range of usability (typical of the recognition heuristics). The extracted factors may be divided into the input factors, and goals.The dimensionality of the input is continuously reduced by a “projection” which drives the new stepsof encoding/perception affecting the PA (weighting/selecting the new information), and extracts theRelevant Features (RF) for the next stages of the decisional processes. AWASS 2012 Edinburgh 10th-16th June 15
    16. 16. Module I - Some Unification Concepts A-Scheme: The Knowledge Vector Pre-attentive activations determine the Input Context Knowledge factorial scores ActivationVector (I) (K) Scheme S1 Factors Example: (1) (1) (1) (1) (1) I1 W1 , W2 . . . , Wn A1 , . . . , AN K1 I2 K2 I3 Scheme S2 K3 LUCKY STRIKE ... (2) (2) (2) (2) (2) ... W1 , W2 . . . , Wn A1 , . . . , AN In KN Scheme activation Factorial scores activation scores Among the activation factors there is also the available time, which contributes (with cognitive cost Relevance Heuristic (R) and conflicts among schemes) to the stress or anxiety: this factor is at the basis of the choice Integrates the external information (EI) between fast&frugal vs "rational" processing of with the “pre attentive activations” (PA) information in order to “choose” if activate a certain A-scheme. An A-scheme so can be characterized in terms of cognitive The conflicts, failures, required times are also used in salience based on its overlapping with the the evaluation/learning phase to promote/devaluate vector (EI*PA) schemes AWASS 2012 Edinburgh 10th-16th June 16
    17. 17. Module I: OverviewThe schemes in module 1 deal with the input factors, while those in module 2propose the goal factor (emotionally related) and when accepted by the goalheuristics these factors may conclude the processing of a given piece ofinformationThe relevance heuristic deals with conflicts among schemes: for instancemore than one scheme may be activated, and the proposed modifications tothe context are in conflict (perceptive dissonance).As  schemes in module 1 one may thing that these schemes have anactivation pattern that has to match the context, and a general score thatdepends on past activity (learning), and that they actively modify thecontext, both the input part and the goal.A possible mechanism of the pattern matching is that the highest the matchwith the context, the faster is the activation of a scheme. AWASS 2012 Edinburgh 10th-16th June 17
    18. 18. Module II: The Conscious knowledge From Cognitive psychology to Probabilistic ReasoningThe theoretical structure of the module II has been developed on the basis of themost relevant models of probabilistic reasoning and social cognition theories, andtries to integrate in a general and psychologically coherent framework their crucialfeatures. Moreover very recent neurophysiological evidences suggest the existenceof different kind of Heuristics (processes) at this stage.Involved cognitive processes Bottom Up processes - e.g. Analogical Mapping of the information Top Down processes - e.g. Reasoning (Decision Making, Problem Solving)Fundamental features Data oriented processes Analogical representation of the Goal/Target Selection/Evaluation and management of the B-Scheme B-Scheme mental simulation and activation AWASS 2012 Edinburgh 10th-16th June 18
    19. 19. Module II: The Conscious knowledge From Cognitive psychology to Probabilistic ReasoningDynamics of MODULE II:Goal Heuristic uses some “components” of K to create the most probable Goal Scheme (GS)(i.e. representation of the goal). This low dimensional scheme has the form of a B-Schemeand is updated with (and updates too) K.Recognition Heuristic integrates the RF coming from module I with GS in order to activatethe most relevant B-Scheme. This could be considered as a continuous and incrementalprocess which is interrupted only by the Solve Heuristic and where a temporary new B-Scheme can be built if required as a linear combination of the previously activated ones(Representativeness, anchoring, availability).Solve Heuristic explicitly explores (frontal activity) the probability of success (distancebetween GS and activated B-Scheme) and the cognitive costs of the activated/created B-Scheme. With a simple function of the previous two arguments the recognition heuristic isstopped (Fast and Frugal, Less is More) when the ratio among goal closeness and cognitivecosts find a local maximum. Alternatively it drives the gathering of new information by themodification (enlargement) of the RF and K. AWASS 2012 Edinburgh 10th-16th June 19
    20. 20. Module II - Some Unification Concepts The goal Scheme Goal Factors: Indicates the expected emotional/physical efforts provided byContext Knowledge(K) the goalK1, K2, K3, ..., KN Goal Goal-Schemes (Gk): Factors Factorial (k) (k) (k) (k) (k) (k) (k) activation scores G1 , G2 . . . , GN F1 , . . . , F N E1 , . . . , E N Extracted G(k) Factors Schemes in module 2 perform actions, and to be accepted they propose emotional goals (solution of the problem) that originate from internal, qualitative goals (bring food to the mouth). In general schemes tends to activate other schemes (mainly by modification of the context), but the actual activation is governed by heuristics, given the available time, cognitive cost, etc. AWASS 2012 Edinburgh 10th-16th June 20
    21. 21. Module II - Some Unification Concepts THE recognition process Knowledge (K) Goal Goal-Schemes (Gk): Factors K1, K2, K3, ..., KN (k) (k) (k) G1 , G2 . . . , GN (k) (k) F1 , . . . , F N (k) (k) E1 , . . . , E N B-Scheme (Bk): Factorial Answer activation scores (h) (h) (h) (h) (h) B1 , B2 . . . , BN F1 , . . . , F N & Cost (h) BRecognition heuristic (RH): the activation of pattern/modification of context in principleis a sort of dynamical process that may end in fixed point or be trapped into a cycle(indecision), but has a structure of an attractor, ... and it takes time to emerge (due tothe action of the recognition heuristics).The first activated schemes are those that have a strong match with the context, and iftime or cognitive resources are  limited the goal heuristic may decide that the goal levelis enough to stop the process.Therefore, for short times, the decision process is essentially a tree, with quite skewedbranches: it is essentially the principle "take the best" (match) of the fast and frugalprocess. AWASS 2012 Edinburgh 10th-16th June 21
    22. 22. Module II - Some Unification Concepts THE solve process Context Knowledge(K) Goal Goal-Schemes (Gk): Factors K1, K2, K3, ..., KN (k) (k) (k) G1 , G2 . . . , GN (k) (k) F1 , . . . , F N (k) (k) E1 , . . . , E N Factorial B-Scheme (Bh): Goal scores (h) (h) (h) (h) (h) Answer B1 , B2 . . . , BN F1 , . . . , FN & CostSolve Heuristic (SH) explores the probability of success and the cognitive costs of theactivated/created B-Scheme.SH stops the Recognition Heuristic (Fast and Frugal, Less is More) when the ratio amonggoal closeness and cognitive costs find a local maximum.Alternatively it drives the gathering of new information by the modification(enlargement) of the Relevant Factors and Knowledge/Context vector. AWASS 2012 Edinburgh 10th-16th June 22
    23. 23. Module III: LearningInside this framework the Learning can be seen as a reinforcement of schemes by means ofcomparisons between expected goals and obtained results. In this sense it can be consideredanalogous to the Hebbian reinforcement assumptions. Nevertheless a fundamental ingredientof learning is the forgetting process, which for instance enables the recognition heuristic andthe fluency heuristic to make better inferences. Involved cognitive processes Bottom Up processes - e.g. Hebbian learning (unconscious learning) Top Down processes - e.g. Social Learning and Mental Simulation Fundamental features Updating and management of the associative and analogical maps (A,B-Schemes) Evaluation of the behaviour related outputs Imitation and Mental Simulation (e.g. internal use of the M-II heuristics) Oblivion processes AWASS 2012 Edinburgh 10th-16th June 23
    24. 24. Module III: LearningDynamics of MODULE III:Evaluation Heuristic compares the External Input with the expected Goal Scheme, andassesses the goodness of the answer (emotional activations).Automatic Learning: Active on A and B-Schemes - Hebbian like reinforcement based onfrequency of occurrences.Observation/Imitation - (Social Learning) Active on B-Scheme - Activation of the sameobserved B-schemes and a consequent Hebbian evolution on the bases of the EvaluationHeuristic result (Symbolic Interactionism theory and Attribution theory).Trial and Error- Active on Scheme B - Evaluation heuristic and Hebbian managing of the B-scheme.Mental Simulation - Induction - Active on Scheme B - New associations or acquaintances canbe represented as new B-Schemes, which are compared with the existing ones by themodule II and then possibly reinforced by the module III (Cognitive dissonance theory). AWASS 2012 Edinburgh 10th-16th June 24
    25. 25. ConclusionThe human cognitive dynamics is based on relatively simple "fast and frugal"procedures, that cooperate in a complex environment.We denote as "schemes" the active procedure that manage information andperform actions, and by "heuristics" the management of schemes: activation,conflict resolution, tuning, learning.Based on time response and imaging techniques it is possible to suggest ahierarchical structure.We propose a unified, tri-partitioned model: a perceptive module I, an actionmodule II and a learning module III.The main connection among schemes is by means of the context frame: a series offactors and of emotional goals (the latter only affecting schemes in module II).Schemes have an associated score, that measures the efficacy of the procedure,the conflicts with other schemes, the cognitive costs. AWASS 2012 Edinburgh 10th-16th June 25 3
    26. 26. ConclusionSchemes in module I are responsible for input processing, extraction of relevantfactors (and of focussing on important pieces of information), and activation of moduleII schemes. The factors contribute to the context frame, which is also the mechanismfor activating other schemes through pattern matching. The only heuristic in module Iis the Relevance Heuristic, responsible of resolving conflicts among schemes.Schemes in module II perform actions and activate other schemes, through thecontext frame. These modules have goals (internal, specific ones and emotional,common ones).There are three heuristics in module 2: the Goal Heuristic that manages the goals, theSolve Heuristic that manages the computational cost of schemes, and the RecognitionHeuristic that eventually activates schemes based on partial matching.Module III is devoted to learning, either by a simple unconscious Hebbianreinforcement based on the score of modules, or on social learning (imitation) andmental simulation (Evaluation Heuristic). AWASS 2012 Edinburgh 10th-16th June 26 3
    27. 27. AWASS 2012 Edinburgh 10th-16th June... and thanks for the attention!

    ×