Cognitive bootstrapping and a priori knowledge
Cognitive bootstrapping and a priori knowledge Cognitive bootstrapping and a priori knowledge
Cognitive bootstrapping and a priori knowledge Andrea Kulakov, Sts Cyril and Methodius University Skopje, Macedonia Georgi Stojanov, American University of Paris Paris, France
- Page 2 and 3: Objectives of the work-package WP5
- Page 4 and 5: Introduction to Curiosity • With
- Page 6 and 7: Schmidhuber (1991) • Schmidhuber
- Page 8 and 9: Weng et al 2001; Huang and Weng 200
- Page 10 and 11: Blank et al (2005) • Blank et al.
- Page 12 and 13: On Innate KnowledgeK • Where does
- Page 14 and 15: Example: Perceiving causation tribu
- Page 16 and 17: How about now This was easier to gu
- Page 18 and 19: Precocial/Altricial • • Precoci
- Page 20 and 21: Biological Nativism: Altricial/Prec
- Page 22 and 23: XPERO’s s first experiments • O
- Page 24 and 25: Types of stimuli • Drives/needs,
- Page 26 and 27: Why innate knowledge • The robot
- Page 28 and 29: Representation of Proto-Objects Fea
- Page 30 and 31: We need to think about architecture
- Page 32 and 33: Deeper notion of functionalism •
- Page 34 and 35: Nervous system of the “cybernetic
- Page 36 and 37: Including alarms (priorities of exe
- Page 38 and 39: H-CogAff cognitive architecture
- Page 40 and 41: CoSy Architecture
- Page 42 and 43: The 4 crucial mechanisms that guide
- Page 44 and 45: simil link Proto-conceptual network
- Page 46 and 47: Building an abstract schema The rel
- Page 48 and 49: Levels of abstraction
<strong>Cognitive</strong> <strong>bootstrapping</strong> <strong>and</strong><br />
a <strong>priori</strong> <strong>knowledge</strong><br />
Andrea Kulakov,<br />
Sts Cyril <strong>and</strong> Methodius University<br />
Skopje, Macedonia<br />
Georgi Stojanov,<br />
American University of Paris<br />
Paris, France
Objectives of the work-package WP5<br />
• To explore the notion of innate <strong>knowledge</strong> <strong>and</strong> cognitive<br />
bootstrap within the main goal of XPERO: learning by<br />
experimentation.<br />
• To investigate the state of the art in computational models of<br />
inborn <strong>knowledge</strong> <strong>and</strong> internal value systems<br />
• overview of existing models of curiosity<br />
• innate <strong>knowledge</strong><br />
• (Guidelines for the XPERO architecture)
WP5 relation to the other WPs<br />
WP 5<br />
Innate Knowledge <strong>and</strong><br />
<strong>Cognitive</strong> Bootstrap<br />
WP 1<br />
Stimulation of<br />
Experiments<br />
WP 4<br />
Gaining Insights<br />
<strong>and</strong> Representing<br />
Knowledge<br />
WP 6<br />
The Experimental<br />
Loop<br />
WP 3<br />
Observation <strong>and</strong><br />
Evaluation of<br />
Experiments<br />
WP 2<br />
Design <strong>and</strong> Execution<br />
of Experiments
Introduction to Curiosity<br />
• With the advent of developmental (epigenetic) robotics we are<br />
witnessing an increased interest in<br />
for autonomous agents <strong>and</strong> especially in the notion of curiosity<br />
witnessing an increased interest in motivational subsystems<br />
• This community is particularly interested in agents that, during their<br />
development, exhibit ever more complex behavior via the much sought<br />
after open-ended ended learning<br />
• Sometimes this type of learning is also referred to as task-<br />
independent or<br />
• Curiosity<br />
or task non-specific<br />
specific.<br />
Curiosity then, in this context, is understood to be the<br />
mechanism that would drive these systems to do<br />
something rather than nothing<br />
• Researchers adopting Piagetian schema constructs point out that<br />
schemas are self-motivated<br />
to be executed
Some implementations of curiosity in<br />
artificial agents
Schmidhuber (1991)<br />
• Schmidhuber (1991) introduces the notion of curiosity<br />
in an otherwise Reinforcement Learning (RL) setup.<br />
S<br />
CUR<br />
model ctrlr<br />
M<br />
• In his agent there are 2 recurrent neural networks<br />
(RNN). The first one models the environment, by<br />
implementing a predictor mechanism:<br />
Situation1-Action<br />
Action-Situation2<br />
<strong>and</strong> the second one (the controller) actually controls<br />
agent’s s behavior (i.e. chooses the next action to be<br />
executed).
Kaplan <strong>and</strong> Oudeyer (2002)<br />
• There are three essential “processes” that interact with each other:<br />
motivation, prediction, , <strong>and</strong> actuation.<br />
• Motivation process is based on three motivational variables:<br />
• predictability (how good is the prediction process in guessing the next<br />
S(t) given the previous (SM(t-1)),<br />
• familiarity (how many times the robot has actually experienced that<br />
particular transition SM(t-1) to S(t)), <strong>and</strong><br />
• stability (how close remains S(t) to its average value).<br />
The reward function is such that the robot gets positive<br />
• The reward function is such that the robot<br />
reinforcement if it maximizes stability motivational variable,<br />
<strong>and</strong> when it maximizes the first derivative of the predictability <strong>and</strong><br />
familiarity motivational variables.<br />
• Apparently this reward policy is a variation of Schmidshuber’s principle.<br />
Kaplan <strong>and</strong> Oudeyer also relate their motivational variables to the t<br />
notions of<br />
novelty <strong>and</strong> curiosity as used by (Huang <strong>and</strong> Weng, , 2002) <strong>and</strong> (Kulakov <strong>and</strong><br />
Stojanov, 2002).
Weng et al 2001; Huang <strong>and</strong> Weng 2002<br />
• Working within the context of a research program called autonomous<br />
mental development (Weng<br />
et al., 2001; Weng, , 2002) Huang <strong>and</strong><br />
Weng (e.g. Huang <strong>and</strong> Weng, , 2002) have implemented a motivational<br />
system in a physical robot called SAIL, which<br />
in a physical robot called SAIL, which rewards the robot for<br />
going into novel situations.<br />
• A novel situation is defined in terms of how different are the current c<br />
sensory inputs from the ones encountered in the past. This pushes s the<br />
robot towards regions where its predictor<br />
errors in guessing the next sensory input, <strong>and</strong>, as expected the robot<br />
indeed improves its performance in environments that are<br />
deterministic <strong>and</strong> learnable.<br />
predictor makes biggest<br />
• The problem arises in probabilistic <strong>and</strong>/or noisy environments<br />
where the robot apparently behaves r<strong>and</strong>omly in order to<br />
maximize the prediction error (<strong>and</strong> the reinforcement with that).
Barto at al, , 2004; Stout et al, , 2005<br />
• Generalize their traditional reinforcement learning approach<br />
(e.g. Sutton <strong>and</strong> Barto, , 1998) by distinguishing between<br />
external reinforcement (usually given by a teacher<br />
or critic) <strong>and</strong> internal reinforcement. . The internal<br />
reinforcement allows for intrinsically motivated<br />
learning which would enable the agent to learn<br />
“[The intrinsic reward system] favors the development<br />
of broad competence rather than being directed to more<br />
specific externally-directed goals. But these skills act as<br />
the “building blocks” out of which an agent can form<br />
solutions to specific problems that arise over its<br />
lifetime.” (Barto et al. 2004)
Blank et al (2005)<br />
• Blank et al. . in (2005) identify three essential notions for<br />
autonomous development in robots:<br />
• abstraction,<br />
• anticipation, <strong>and</strong><br />
• self-motivation.<br />
• The self-motivation subsystem:<br />
“[…] indicates to the system how “comfortable” it is in the given<br />
environment. If it is too comfortable, it becomes bored, <strong>and</strong> takes<br />
measures to move the robot into more interesting areas.<br />
Conversely, if the environment is chaotic, it becomes over-excited<br />
<strong>and</strong> attempts to return to more stable <strong>and</strong> well known areas.”<br />
• They present the initial results on a simulated agent that solves s the<br />
navigational problem.
A spectrum of competences<br />
• • Every organism is a mixture of both kinds of capabilities:<br />
• pre-configured — constructed (meta-configured)<br />
tribute to A. Sloman<br />
• • Not all of the pre-configured capabilities are manifested at birth –<br />
many are ‘time-bombs’ (e.g. waiting for the season to hibernate, or migrate).<br />
• • Architectures for the more advanced species can do many things<br />
that are not directly biologically useful, , but may provide reusable<br />
information: including (possibly dangerous) exploration of a space of possibilities<br />
lities.<br />
• • Architectures can change over time.<br />
• • Ontologies used can change over time.<br />
• • Forms of representation used can change over time.
On Innate KnowledgeK<br />
• Where does <strong>knowledge</strong> come form<br />
• Nativists<br />
• Plato<br />
• Chomsky<br />
• Nurturists<br />
• Behaviorists (Brooks, …)<br />
• <strong>Cognitive</strong> Developmentalist<br />
• Related projects<br />
Related projects<br />
• The phylogenetic abilities (the set of sensori-motor circuits) are<br />
predetermined by the chosen ontology of innate concepts.<br />
RobotCub people decided on the following list of innate concepts:<br />
objects, numbers, space <strong>and</strong> people.<br />
• innate skills of the COGNIRON robots
Innate skills of the COGNIRON robots
Example: Perceiving causation<br />
tribute to A. Sloman<br />
• Our ability to perceive moving structures, <strong>and</strong> our<br />
meta-level ability to think about what we perceive,<br />
is intimately bound up with perception of causation<br />
<strong>and</strong> affordances.<br />
• Sometimes the causal relations are inherent in what<br />
is seen<br />
• Sometimes they involve invisible (hypothesised)<br />
structures <strong>and</strong> processes
What is the right gear going to do if the left<br />
one is turned clockwise<br />
We do not know the functionality (the mechanism) inside the box!
How about now<br />
This was easier to guess, but …
Why not something like this<br />
… we assumed the rigidity of the materials from which the gears are made!
Precocial/Altricial<br />
• • Precocial<br />
• Some animals are born highly competent: deer, chickens, etc.<br />
tribute to A. Sloman<br />
• • Altricial<br />
• Some animals are born underdeveloped <strong>and</strong> highly incompetent, but adult<br />
forms can do things precocial species cannot, e.g. hunting mammals, nestbuilding<br />
birds, primates, humans.<br />
• • Even altricial species have some precocial skills or tendencies<br />
• e.g. sucking, stimulating parents to feed, <strong>and</strong> some ‘delayed’ precocial skills,<br />
like sexual maturation.<br />
• • Architectures <strong>and</strong> competences may be pre-formed in precocial<br />
species, but slightly adaptable, e.g. by reinforcement learning.<br />
• to contrast learning a language, or learning to program computers<br />
• • Altricial species may be using sophisticated architecture-growing<br />
mechanisms doing far more than varying weights (etc.), when they look<br />
incompetent.<br />
• By collecting chunks of information about affordances provided by the<br />
environment <strong>and</strong> by their bodies — initially these affordances are stored,<br />
then later are recombined <strong>and</strong> used.
Biological <strong>bootstrapping</strong> mechanisms<br />
tribute to A. Sloman<br />
• • There are some species whose needs cannot be served by genetically<br />
ly<br />
determined (preconfigured) competences based on pre-designed<br />
architectures, forms of representation, ontologies, mechanisms, <strong>and</strong> stores<br />
of information about how to act so as to meet biological needs.<br />
• • Evolution have ‘discovered’ that it is possible instead to provide a<br />
powerful meta-level <strong>bootstrapping</strong> mechanism for ‘meta-configured’<br />
species:<br />
• a mechanism without specific information about things that exist in the<br />
environment (apart from very general features such as that it includes spatio-temporal structures <strong>and</strong><br />
processes, causal connections, <strong>and</strong> opportunities to act <strong>and</strong> learn, <strong>and</strong> that the neonate has a body that is<br />
immersed in that environment)<br />
• with specific information about types of: things to try doing, things to observe,<br />
things to store<br />
• with specific information about how to combine the things done <strong>and</strong> keep<br />
records of things perceived into ever larger <strong>and</strong> more complex reusable<br />
structures,<br />
• including a continually extendable ability to run simulations that can be used<br />
for planning, predicting <strong>and</strong> reasoning.
Biological Nativism: Altricial/Precocial<br />
tradeoffs<br />
tribute to A. Sloman<br />
• • Evolution ‘discovered’ that for certain species which need to adapt relatively<br />
quickly to changing environmental pressures, a kind of learning mechanism is<br />
possible which combines previous <strong>knowledge</strong> <strong>and</strong> allows much faster <strong>and</strong> richer<br />
learning than is possible in systems that merely adjust probabilities on the basis of<br />
observed evidence (statistical data).<br />
• • The altricial species learn a great deal about the environment after birth <strong>and</strong> in<br />
some cases are able rapidly to develop capabilities none of their r ancestors had<br />
• like young children playing with computer games.<br />
• • This uses an information-processing architecture which starts off with a collection<br />
of primitive perceptual <strong>and</strong> action competences, but also with a mechanism for<br />
extending those competences by ‘syntactic’ composition<br />
• as a result of play <strong>and</strong> exploration, which is done for its own sake, not to meet<br />
other biological needs (food, protection from hurt, warmth, etc.)<br />
• • The meta-level features of the mechanism <strong>and</strong> the initial competences are<br />
genetically determined, but the kinds of composite competences that t<br />
are built are<br />
largely a function of the environment.<br />
• • This requires forms of learning that are not simply adjustments of probabilities,<br />
but involve continual creation of new useful structures, exp<strong>and</strong>ing ng the ontology<br />
used.
The developmental architecture<br />
tribute to A. Sloman<br />
• • There is an important sub-class of animals in which competences are not<br />
all pre-configured, whose development makes use of:<br />
• Genetically determined primitive actions, perceptual capabilities <strong>and</strong><br />
representations,<br />
• Genetically determined play/exploration mechanisms which drive<br />
learning that extends those actions, etc., using abilities to chunk,<br />
recombine <strong>and</strong> store<br />
• new more complex action fragments<br />
• new more complex perceptual structures<br />
• new more complex goals<br />
• Creating new ontologies, theories, competences (cognitive <strong>and</strong> behavioural)<br />
• i.e. new more complex thinking resources,<br />
• • Not restricted to somatic sensorimotor ontologies.<br />
• • Thereby extending abilities to search in a space built on larger chunks:<br />
solving ever more complex problems quickly.<br />
• unlike most statistical forms of learning<br />
• • For AI systems this will require us to discover new architectures s <strong>and</strong><br />
learning mechanisms.
XPERO’s s first experiments<br />
• One robot <strong>and</strong> one object<br />
• Motivation, internal value system, experimental<br />
stimuli: : implicit/nonexistant<br />
• The notion of object is implicit: : the designer<br />
chooses the set of learning attributes/features<br />
(distance, angle…)<br />
• Motor comm<strong>and</strong>s generated by the designer: tele-<br />
operated robot
Eventually the embodied robot<br />
• Will need motivation <strong>and</strong> internal value system:<br />
•Why would the robot do anything<br />
• Self-preservation<br />
• Curiosity<br />
• Explaining unexpected phenomenon<br />
•What would the robot do<br />
• Stochastic experimentation<br />
• “Planned” experimentation
Types of stimuli<br />
• Drives/needs, , a stimulus which arises either extrinsically<br />
or intrinsically to secure the survival <strong>and</strong> integrity of the<br />
embodied agent, , e.g. obtain food <strong>and</strong> energy, procure<br />
shelter, liberate from captivity or from an emergency<br />
situation.<br />
• Curiosity, , the innate interest to find <strong>and</strong> explore the<br />
unknown, may it be unknown physical space, unknown<br />
objects, unknown functions <strong>and</strong> properties of objects or<br />
unknown own capabilities<br />
• Novelty triggering hypotheses formation about a<br />
physical phenomenon which has emerged from a current<br />
activity of the embodied agent, e.g. during the execution of<br />
a task;
Stochastic vs. planned experiments<br />
• Stochastic experimentation occurs during the<br />
cognitive bootstrap; the chosen elementary actions<br />
(contingent on the embodiment) <strong>and</strong> the inborn<br />
proto-objects objects provide the robot with the initial<br />
ontology (what type of objects there are)<br />
• Planned experiments to gain insights about<br />
properties <strong>and</strong> relations among objects
Why innate <strong>knowledge</strong><br />
• The robot needs a mechanism for creating the basic<br />
ontology<br />
• The notion of proto-object<br />
object will speed up the<br />
development<br />
• As we are not necessarily doing modeling of<br />
cognitive development we can introduce what we<br />
deem fit (logic, self-preserving behaviors…)
Innate <strong>knowledge</strong><br />
• Self-preserving reflex behaviors<br />
• Proto-objects<br />
objects<br />
• Innate gestalt principles (possible point of<br />
interest in the sensory input)<br />
• Internal value system<br />
• Logical inference rules
Representation of Proto-Objects<br />
Feature1:<br />
Symbolic<br />
name<br />
Feature2:<br />
FeatureN:<br />
affordancy1<br />
affordancN<br />
activation
Relations between objects<br />
Feature1:<br />
Feature2:<br />
FeatureN:<br />
Sym<br />
Name<br />
affordancy1 affordancyN<br />
activation<br />
Instance of<br />
A<br />
Feature1:<br />
Feature2:<br />
FeatureN:<br />
Instance of<br />
B<br />
Feature1:<br />
Feature2:<br />
FeatureN:<br />
affordancy1affordancyN<br />
activation<br />
affordancy1 affordancN<br />
activation<br />
C<br />
Feature1:<br />
Feature2:<br />
FeatureN:<br />
D<br />
Feature1:<br />
Feature2:<br />
FeatureN:<br />
affordancy1<br />
affordancyN<br />
affordancy1 affordancyN<br />
activation<br />
activation<br />
Later we will summarize the 4 crucial mechanisms that guide agent’s behavior
We need to think about architectures<br />
tribute to A. Sloman<br />
• • The sort of system we are discussing has many components doing many m<br />
different things in parallel.<br />
• • Putting the pieces together in a working architecture is a non-<br />
trivial task for engineers <strong>and</strong> for scientists attempting to produce explanatory models<br />
scientists attempting to produce explanatory models.<br />
• • We need good theories about the space of possible architectures,<br />
<strong>and</strong> good theories about particular architectures in that space in order to explain the wide w<br />
variety of biological phenomena <strong>and</strong> in order to underst<strong>and</strong> the development of humans, since<br />
we are not born with a fully fledged architecture: their architecture grows in ways that may<br />
partly replicate some of our evolutionary history but will be much influenced by our culture <strong>and</strong><br />
physical environment.<br />
• • Different aspects of motivation <strong>and</strong> emotion relate to different<br />
architectural layers with different competences.
Functionalism Why architectures,<br />
when we can use FSM<br />
tribute to A. Sloman<br />
• • Functionalism is one kind of attempt<br />
to underst<strong>and</strong> the notion of virtual<br />
machine, in terms of states defined by<br />
a state-transition table (Finite-State-<br />
Machines - FSM).<br />
• • This is how many people think of<br />
functionalism: there’s a total state<br />
which affects input/output<br />
contingencies, <strong>and</strong> each possible state<br />
can be defined by how inputs<br />
determine next state <strong>and</strong> outputs.<br />
State(t+1) = f(State(t), Act(t), Observ(t))
Deeper notion<br />
of functionalism<br />
• Instead of a single (atomic) state which<br />
switches when some input is received, a<br />
virtual machine can include many subsystems<br />
with their own states <strong>and</strong> state<br />
transitions going on concurrently, some<br />
of them providing inputs to others.<br />
• The different states may change on<br />
different time scales: some change very<br />
rapidly others very slowly, if at all.<br />
• They can vary in their granularity: some<br />
sub-systems may be able to be only in<br />
one of a few states, whereas others can<br />
switch between vast numbers of<br />
possible states (like a computer’s<br />
virtual memory).<br />
• Some may change continuously, others<br />
only in discrete steps. Some subprocesses<br />
may be directly connected to<br />
sensors <strong>and</strong> effectors, whereas others<br />
have no direct connections to inputs<br />
<strong>and</strong> outputs <strong>and</strong> may only be affected<br />
very indirectly by sensors or affect<br />
motors only very indirectly (if at all!).<br />
tribute to A. Sloman
Under development<br />
tribute to A. Sloman<br />
• A TAXONOMY OF TYPES OF ARCHITECTURE, based on<br />
the analysis of:<br />
• Requirements for architectures,<br />
• Designs for architectures,<br />
• Components of architectures<br />
• Varieties of information structures<br />
• Varieties of mechanisms<br />
• Kinds of control systems<br />
• Ways of assembling components<br />
• How architectures can develop,<br />
• Tools for exploring <strong>and</strong> experimenting with<br />
architectures
Nervous system of the<br />
“cybernetic animal”<br />
Nerve net<br />
Receptors<br />
Effectors<br />
Environment<br />
adapted from V. Turchin, “The Phenomenon of Science”, Columbia University Press, 1977
The Reactive Architecture
Including alarms (<strong>priori</strong>ties of execution)
More developed architecture
H-CogAff<br />
cognitive architecture
Interactivist cognitive architecture
CoSy Architecture
Proposed XPERO general architecture
The 4 crucial mechanisms that guide agent’s<br />
behavior / development (Kulakov, Stojanov 2002)<br />
• Abstraction mechanism that provides chunking of the sensory<br />
that provides chunking of the sensory-<br />
motor flux from highly dimensional input via proto-objects objects to symbolic compact<br />
descriptions of the objects in the environments. This will be happening during<br />
the cognitive bootstrap phase using the stochastic experimentation. Abstraction<br />
mechanism will enable the agent to deal with more <strong>and</strong> more complex situations<br />
with the same or less cognitive effort;<br />
• Planning mechanism used during deliberate experimentation. It<br />
combines various previous experiences into new <strong>knowledge</strong>, for example by<br />
analogy-making;<br />
Mechanism that provides emergence of<br />
• Mechanism that provides emergence of more complex inner<br />
value <strong>and</strong> motivational systems according to which new<br />
experiences are judged, foreseen <strong>and</strong> executed;<br />
• Socialization/communication mechanism that enables the<br />
agent to interpret in a special way inputs coming from other agents (possibly<br />
humans) <strong>and</strong> to provide translation of the newly acquired <strong>knowledge</strong> into human<br />
underst<strong>and</strong>able form;
A schema<br />
schema links with weights denoting<br />
the reliability of the schema<br />
p1<br />
W1<br />
W2<br />
p2<br />
Condition Percept Schema Node Expectation Percept
simil link<br />
Proto-conceptual network<br />
abstract schema<br />
percept node<br />
schema node<br />
schema link<br />
type/token link
Building an abstract schema<br />
Condition<br />
Percept<br />
Expectation<br />
Percept<br />
Action Sequence<br />
or<br />
Abstract Schema<br />
If long enough sequence of reliable schemas<br />
is detected (3+ actions), then a new Abstract<br />
schema is created with the same Condition<br />
Percept as the first schema <strong>and</strong> the same<br />
Expectation Percept as the last schema<br />
in the sequence.<br />
The Action Sequence of the Abstract<br />
Schema is left undefined, but<br />
type/token links are created between<br />
the Abstract Schema <strong>and</strong> the<br />
underlying more<br />
concrete schemas.
Building an abstract schema<br />
The reliability of the abstract schema<br />
takes the value of the lowest reliable<br />
schema at the level below (the weakest<br />
link), while its level of abstraction equals<br />
the level of the most abstract schema<br />
between its constituents, plus one<br />
If the Expectation Percept<br />
matches the Current Percept<br />
derived from the Current Sensory<br />
Input, then the Reliability of the<br />
Schema is increased; otherwise it<br />
is lowered
Part of the internal representation ( (proto-<br />
conceptual network)
Levels of abstraction
Important<br />
• Internal value system <strong>and</strong> the previous <strong>knowledge</strong><br />
influence the perception, the observation <strong>and</strong> the<br />
evaluation of an ongoing experiment (WP1 <strong>and</strong><br />
WP3), as well as the planning <strong>and</strong> decision making<br />
processes (WP2)<br />
• It is very important to know the architecture of the<br />
whole system (what is learnt, how it is learnt, how<br />
it is represented, which part makes the decision<br />
what to do next <strong>and</strong> how, etc.)