Presentation on theme: "Design principles for adaptive self-organizing systems"— Presentation transcript:
1Design principles for adaptive self-organizing systems Finding Fluid Form SymposiumUniversity of BrightonDecember 9-10, 2005Design principles for adaptive self-organizing systemsPeter CarianiDepartment of PhysiologyTufts Medical SchoolBoston
2My trajectoryOrganismic biology MIT mid 1970s) Biological cybernetics & epistemology (1980s) Biological alternatives to symbolic AI Howard Pattee, Systems Science, SUNY-Binghamton Temporal coding of pitch & timbre (1990s) Auditory neurophysiology, neurocomputation How is information represented in brains? Commonalities of coding across modality & phyla Neural timing nets for temporal processing Auditory scene analysis Possibilities inherent in time codes Temporal alternatives to connectionism signal multiplexing; adaptive signal creation broadcast
3Evolution of ideasElaboration of structures & functions over time in biological, social, and technological realms, What makes new functions possible (functional emergence)? Can we put these principles to work for us? Is structural complexification by itself sufficient? (No) Notions of function & functional emergence are needed. What kinds of functions? Sensing, effecting, coordinating Is pure computation on symbols sufficient? (No) How are brains/minds capable of open-ended creativity? Neural codes, temporal codes, timing nets Neural coding of pitch in the auditory system Rethinking the architecture of the brain: Temporal alternatives to connectionism Adaptive signal creation & multiplexing, Broadcast coordinative strategies
5An example Exhaustive description Limited description All permutations ofsingle digitsconsisting of 6 tokensAll permutations of6 arbitrarily defined objectsOne well-defined sethaving 610 permutationsBOUNDEDIll-defined number of sets, each w. 610 permutationsUNBOUNDED
6Describing the world: Two perspectives Omniscent“God’s eye view”Postulational,ontologicalanalytical modePerspective ofthe limited observerepistemologicalempirical modeAppearance of newstructures over timeViolations of expectations“Surprise”
7Well-defined vs. ill-defined realms Exhaustive descriptionGod’s eye viewLimited descriptionLimited observerSystem-environment aswell-defined realmEnvironment asill-defined realmDescription isdependent on set of observables(environment has as many properties as one can measure)Description ofall-possibleorganism-environmentrelationsCLOSED WORLD ASSUMPTIONOPEN WORLD ASSUMPTIONNo fundamental novelty is possibleAll novelty is combinatoricCombinatoric andCreative emergence
8New features CREATING A NEW OBSERVABLE ADDS A NEW PRIMITIVE THAT INCREASES THE EFFECTIVE DIMENSIONALITY OF THE SYSTEM
9PhilosophyOntology Aristotelian hylomorphism Material substrate that exists independently of us, yet whose form is largely ill-defined, incompletely known Organization is embedded in material system (e.g. mind is the organization of the nervous system) Conscious awareness requires a particular kind of regenerative informational organization embedded in a material system (cybernetic functionalism) Aristotle's Causes: Multiple complementary modes of explanation that answer different kinds of questions
11PhilosophyEpistemology Pragmatism (truth of a model related to its purpose) Perspective of the limited observer Relativism: different observational frames & purposes Analytical, empirical and pragmatic truths Analytic: truths of convention (non-material truths, finist mathematics) Empirical: truths of measurement, observation (science) Pragmatic: truths of efficacy & aesthetics (engineering, art) Constructivism & epistemic autonomy: by semi-freely choosing our own observables & concepts, we construct ourselves (for better or worse)
12adaptive, self-organizing systems Design principles foradaptive, self-organizing systemsWe are interested in designing & fabricating systems that autonomously organize themselves to elaborate structures & improve functions in response to challenges of their environments in ways that are meaningful and useful to us and/or them
13Design principles for adaptive, self-organizing systems Richness of material possibility (e.g. polymeric combinatorics) + Ability to steer & stabilize structure (feedback to structure: sensors, coordination mechanisms, effectors) + Means to interact w. material world (sensing, action = "situatedness", semantics) + Means to evaluate actions re: purposes (goal-laden representations, "intentionality") => Material system capable of adaptive, elaboration & improvement of informational functions
14Design principles for adaptive, self-organizing systems Richness of material possibility (need polymers, replicated aperiodic structure, Schrodinger's aperiodic crystal, analog dynamics, ill-defined interactions) Ability to steer & stabilize structure (need controls on self-production of internal structure, enzymes) Means to interact w. material world (Need sensors, effectors, neural nets) Means to evaluate actions re: purposes (Need natural selection or internal goal states, limbic system)
15Vibratory dynamics of matter Cymatics:Bringing Matter to Lifewith SoundHans JennyRichness of material possibilityComplexity is easySteerable complexity is hard
16Two phases in creative learning processes Design principles for adaptive, self-organizing systemsVARIATION + SELECTION + INHERITANCE => ADAPTATIONMaterial possibility+ Steer, stabilize, specify, inherit + Sensorimotor interaction + Evaluation => ASOSTwo phases in creative learning processesExpansive phase: generation of possibilityRealm of free & open creatione.g. scientific imagination and hypothesis creationContractive phase: selection of best possibilitiesRealm of clarity & rigorous evaluatione.g. hypothesis testing (clarity, removal of ambiguity)
17Analog dynamics and discrete symbols We will also argue that one almost inevitably needsmixed analog-digital systems for complex systems:i.e. systems w. analog dynamics constrained by digital states ("symbols")for reliable replication of functionfor inheritability of adaptive improvementsAnalog and digital are complementary modes of descriptionanalog descriptions - continuous differential equationsdigital descriptions - discrete states & ST rules/probabilitiesDigital states or discrete symbols are ultrastable basins of attraction
18Different theoretical approaches to understanding brains and their functions DynamicalsystemsapproachesNeuralinformationprocessingSymbol-processingdifferential growthhomeostasisanalogrepresentationsprocessingstates & switchesbranchingdiscrete
23Self-conscious description of the modeling process: Hertzian modeling relation: measurement & computation
24The choice of observables Finding the variablesThe would-be model maker is now in the extremely common situation of facing some incompletely defined "system," that he proposes to study through a study of "its variables." Then comes the problem: of the infinity of variables available in this universe, which subset shall he take? What methods can he use for selecting them?W. Ross Ashby, "Analysis of the system to be modeled" in: The Process of Model-Building in the Behavioral Sciences, Ohio State Press, pp ; reprinted in Conant, ed. Mechanisms of Intelligence
25Choice of primitive features for classifiers The choice of observables - analogous problemsChoice of primitive features for classifiersEvolution of sensory organs in organismsChoice of sensors for robots
26Semiotics of adaptive devices Feedback to stateFeedback to structurealters functionalities
30Pure computation (state-determined system, no independent informational transactions w. environment)
31Fixed robotic deviceFixed sensors,coordinators,and effectors;Purely reactiveand drivenby its inputs;Incapable oflearning
32Computationally adaptive device Trainable machinesNeural networksAdaptive classifiersGenetic algorithmsRobots w. adaptive programsCapable of learningnew percept-actionmappings (classifications)
33Some observations about adaptability Whatever functionalities are fixed, the designer must specifyworks for well-defined problems & solutionsadvantage: predictable, reliable behaviordrawback: problems of specificationWhatever is made adaptive must undergo a learning phaseneeded for ill-defined problems & solutionssome unpredictability of solutions foundcreative behavior!the more autonomy, the more potentially creativeConsequently, there are tradeoffs betweenadaptability & efficiencyautonomy/creativity & control/predictability
34Evolution/adaptive construction of new sensors sensory evolutionimmune systemsperceptual learningcapable of learningnew perceptualcategoriesnew featureprimitives(new observables)
35Epistemic autonomy• When a system can choose its own categories –through which it perceives and acts on the world –that system achieves some limited degree of epistemic autonomy.• A rudimentary electrochemical device was built by cyberneticist Gordon Pask in 1958 that grew its own sensors to create its own “relevance criteria.”
36"With this ability to make or select proper filters on its inputs, such a device explains the central problem of epistemology. The riddles of stimulus equivalence or of local circuit action in the brain remain only as parochial problems." Warren McCulloch, preface ,Gordon Pask (1961) An Approach to Cybernetics.
37Principles of action/use 1. Front-ends for trainable classifiersUseful in ill-defined situations where one does not a priori know what features are adequate to effect a classification2. Adaptive, self-organizing sensorsGrow structures over analog-VLSI electrode arrays in order to sense new aspects of the world. Use biochemical and/or biological systems coupled to an electrode array3. Materially-based generator of new behaviors (adaptive pattern-generators)Similar steerable, ill-defined systems could be used to generate new patterns (sound, images) in an open-ended way that is not at all obvious to the observer/controller4. Epistemic autonomyDevice chooses how it will be connected to the outside world; what aspects of the material world (categories) are relevant to it. (Symbol grounding, frame problem)
38Feedback to state vs. feedback to structure A thermostat is limited in the information that it can gain from its environment by the fixed nature of its sensors. It has feedback to state, but not feedback to structure. The amount of information that such a system can extract from its environment is finite at any time, and bounded by its fixed structure.A system capable of sensory evolution or perceptual learning has the ability to change its relation to its environs. Such a system has an open-ended set of observational primitives. It has both feedback to state and feedback to structure. The amount of information that such a system can extract from its environment is finite at any time, but unbounded. Such a system is open-ended.
39Analog dynamics without inheritable constraint (Hans Jenny)
41Inheritable construction analog dynamics constrained & selected by discrete symbols Purely analog adaptivesystem must be trainedeach generationGenetic algorithm +Pattern grammar forguiding constructionconstrained searchSymbolically-encodedmemory permits results ofan optimization process tobe passed to subsequentgenerations
44Relation to Ashby's homeostat Analogsensor/controllerUniselector25x25x25x25 = 390k construction possibilities -> variety of the control system, unconstrained searchEvaluation of abilityto control inputs
45Relation to Ashby's homeostat Analogsensor/controllerUniselector25x25x25x25 = 390k construction possibilities -> variety of the control system, unconstrained searchEvaluation of abilityto control inputs
46(ill-defined structure) Ashby's homeostatAnalogcontroller(ill-defined structure)Adaptive analogcontrollerStructure ofparticular controllersis unknown to designerRequisite variety forcontrol is the number ofalternative controllersavailable25x25x25 = 390,625Uniselectorevaluate(in bounds?)Environment
47The homeostat & the brain A few cybernetics-inspired accounts of brain functionSommerhoff (1974) Logic of the Living BrainKlopf, The Selfish NeuronArbib, The Metaphorical BrainMost successful neuroscientific application of cybernetics:W.Reichardt's analysis of fly optomotor loopThe homeostat never caught on as a brain metaphorSome possible reasons:Homeostats never were cast in terms of neural netsNo obvious digital uniselector function in the brainPredominance of problems of pattern recognition andformulation of coherent action over simple problems of internal regulation
48The brain as an adaptive self-organizing system Ideas that flow from cybernetics and theoretical biology:Brains as signal self-production systemsrelated to reverberant loops (a la Lorente, Lashley, Hebb, McCulloch, Pitts & many others)2) Brains as pattern-resonance systemsrelated to Lashley, Hebb, many others3) Brains as multiplexed signaling and storage systemsholographic paradigms, Longuet-Higgins, Pribram,John4) Brains as mass-dynamics, broadcast systems5) Brains as communications nets that create new signals6) Brains as temporally-coded pulse pattern systemsI believe all this is possible using temporal pattern codes.
53Autopoiesis and autocatalysis Life is built upon cycles of self-production
54Brain function may be based on self-productions of spike patterns Hebbian reverberant eigenstates and regenerative temporal patternsMcCulloch & Pitts (1943) Nets with circles render activity independentof time and semi-autonomous re: the environmentvon Foerster (1948) brain eigenstates as a form of ST memory
55Why the mind is in the head Warren McCulloch L.A. Jeffress, ed. Cerebral Mechanisms of Behavior(The Hixon Symposium, Wiley, 1951, reprinted inEmbodiments of Mind, MIT, 1965, concluding lines)This brings us back to what I believe is the answer to the question:Why is the mind in the head?Because there, and only there, are hosts of possible connections to be formed as time and circumstance demand. Each new connection serves to set the stage for others yet to come and better fitted to adapt us to the world, for through the cortex pass the greatest inverse feedbacks whose function is the purposive life of the human intellect.The joy of creating ideals, new and eternal,in and of a world, old and temporal,robots have it not.For this my Mother bore me.
56The brain as a self-regenerating pattern-resonance system
57Tuning in nervous systems Minds as pattern-resonances The same [resonance] is true of all bodies which can yield notes. Tumblers resound when a piano is played, on the striking of certain notes, and so do window panes. Nor is the phenomenon without analogy in different provinces. Take a dog that answers to the name "Nero." He lies under your table. You speak of Domitian, Vespasian, and Marcus Aurelius Antonius, you call upon all the Roman Emperors that occur to you, but the dog does not stir, although a slight tremor of his ear tells you of a faint response of his consciousness. But the moment you call "Nero" he jumps joyfully towards you. The tuning fork is like your dog. It answers to the name A. Ernst Mach, Popular Lectures, “The fibers of Corti” c. 1865
58Pattern resonances: neural assemblies emitting annotative tag signals that elaborate a regenerating signal pattern
59Phase-locking in visual neurons Temporal pattern codesPhase-locking in visual neurons(Horseshoe crab ommatidium, 5-15 Hz flashes)Phase-locking in auditoryl neuronsCat auditory nerve fibers, 250 Hz toneJavel,
60Phase-locking in auditory nerve fibers 250 Hz tone
62Phase-locking of discharges in the auditory nerve Cat, 60 dB SPL
63Temporal coding in the auditory nerve Work with Bertrand DelgutteCariani & Delgutte (1996)Dial-anesthetized cats.100 presentations/fiber60 dB SPLPopulation-interval distributions are compiled by summing together intervals from all auditory nerve fibers.The most common intervals present in the auditory nerve are invariably related to the pitches heard at the fundamentals of harmonic complexes.
70Potential advantages of temporal pattern pulse codes & timing nets Multiplexed signal transmissionOrthogonality of patterns; less interferenceFlexible multimodal integrationEncoding of signal identity in itself (logical type)Liberate signals from wiresBroadcast of signals + selective receptionNonlocal computational operationsMass action (statistical representations)Open-ended creation of new signal primitives
71Temporal expectancies in perception Music, brain, and timeIn the image of the digital computer, we conceptualize brains as distributed logic machines.However, temporal correlation machines may prove to be a better metaphor.Temporal expectancies in perceptionTemporal patterning of body processesTemporal structure of movementTemporal expectations and reward structure(dopamine system, conditioning)Temporal memory tracesMusic may have the profound effects that it does because 1) it directly impresses its temporal structure on the activity of many neuronal populations, and2) the neural codes & computations underlying experience are inherently temporal.Andy Partridge, xtc
72ConclusionsDesign principles for self-organizing systemsStructural complexity alone is not sufficient Pure computation alone is not sufficient Requisites Sensors & effectors Mixed digital-analog design Feedback to structure, self-production Inheritable, replicable (digital) plans Combinatorics of digital strings Rich analog, ill-defined dynamics Goal states and steering/selection mechanisms Possibility of brain as temporally-coded self- organizing system
80Codes are defined in terms of their functional roles What spike train messageshave the same meanings?(functional equivalence classes)What constitutesa differencethat makes a difference?Temporal codes are neuralcodes in which timings ofspikes relative to each otherare essential to theirinterpretation.
83Phase-locking of an LGN unit to a drifting sinusoidal grating TemporalmodulationfrequencyPST HistogramsInterval Histograms4 Hz8 Hz16 Hz32 Hz64 Hz
84Adaptive systems Adaptation ~ adjustment Sensing ~ measurement • Depending upon the self-modification process,adaptive systems change in different ways.• They become tuned to their environments,on the percepton the action sideinternally: anticipating events, forecasting effects• New sensors create new linkages with the external world new perceptual primitivesnew observablesnew modes of adjustment• New effectors create new modes of action
87Frequency ranges of (tonal) musical instruments 10k86543210.50.25Frequency ranges of (tonal) musical instruments27 Hz4 kHz262Hz440880110
88Measurement and tuning mediates interactions with external worldpermittingbehavior contingent upon perceptionAdaptive systems that create their own measurementsare possible (we may be such systems)Tuninginvolves adjustment of internal relations to external relations, i.e. adaptive resonanceIt is possible to envision brains and mindsas resonant systems that operate on patternsrather than coupled via energetic relations
89Areas of self-modifying media Self-modifying computersCoevolution between humans and computersEmergent human-machine couplingsPask’s Conversation theoryComputers need means of independentlyaccessing the world and creating theirown concepts (epistemic autonomy)Self-organizing materialsElectrochemicalFerromagneticBiological-silicon interfacesIntelligent materialsMixed digital-analog feedback systems
90Phase-locking of an LGN unit to a drifting sinusoidal grating TemporalmodulationfrequencyPST HistogramsInterval Histograms4 Hz8 Hz16 Hz32 Hz64 Hz
94Build-up and separation of two auditory objects Two vowels with different fundamental frequencies (F0s) are added togetherand passed through the simple recurrent timing net. The two patterns build upIn the delay loops that have recurrence times that correspond to their periods.Period = 10 msVowel [ae]F0 = 100 HzVowel [er]F0 = 125 HzPeriod = 8 ms
95Contingent vs. logically-necessary “truths” Sensing vs computingContingent vs. logically-necessary “truths”Methodological issues: What distinguishes sensing from other kinds of informational operations? A sensing process must be contingent, it must have two or more possible outcomes to reduce uncertainty, whereas A computation (formal operation) must be logically-determined, it must always produce the same outcome given the same initial state
96Computers and brains•Digital computers presently are capable of recombination-based creativity, but do not presently create new primitives for themselves.• Brains, on the other hand, are self-modifying systems with rich analog dynamics that can serve as substrates for formation of new informational primitives.• Contemplation of self-modifying systems is essential if we are to construct artificial systems that can create meaning for themselves.• We need such systems when problems are ill-defined, or when we desire open-ended creative possibilities.
98Overview I: Measurement in adaptive systems • We discuss the semiotics and functional organization of different adaptive systems.• Adaptive systems reorganize their internal structure in order to improve their performance.• We consider how systems with sensors, effectors, and coordinative faculties can adaptively modify their internal structures and functions.• We consider how this adaptivity leads to emergent functions and behaviors.
99Overview IV: Creativity, autonomy, and specification • Creativity has two levels:Recombination of existing primitivesDe novo creation of new kinds of primitives• Inherent tradeoffs:Specifiability vs. autonomyPredictability/reliability vs. creativity
102Conceptions of “emergence” • Appearance of new structures, functions, behaviors• Novelty that was not predictable from what came beforeVarieties• Structural emergence (appearance of new structures, org. levels)• Computational emergence (unexpected results)• Thermodynamic emergence (dissipative systems)• Functional emergence (flight, color vision)• Emergence-relative-to-a-model (perspectivist, operationalist)
103Methodological issues How can we identify the existence of information processing operations in artificial and natural systems?How can we distinguish measurement, computation, and effector operations from each other in an unknown material system?How can we detect changes in these functionalities, such that we know that our devices or organisms have modified them adaptively?We need operational distinctions.We need to be able to parse a state-transition graph.
109Opening up the sensory interface: Break-out strategies for creating new observables 1) construction of new sensors2) modification of existing sensors3) interposition of sensory prostheses4) active measurements5) creation of new internal sensors
110Prosthesis: augmentation of functionalities All technology is prosthesis.
111Operational states and procedures in a scientific model Explicate realmof symbols(well-defined)Implicate realmof material process(ill-defined)