The Jean System
Yu-Han Chang, Clayton T. Morrison,
Wesley Kerr, Aram Galstyan,
Paul R. Cohen, Carole Beal
USC Information Sciences Institute
4676 Admiralty Way
Marina del Rey, CA 90292
{ychang,clayton,galstyan,wkerr,cohen,beal}@isi.edu
Robert St. Amant
North Carolina State University
Campus Box 8206
2268 Engineering Building II
890 Oval Drive
Raleigh, NC 27695
stamant@csc.ncsu.edu
Tim Oates
University of Maryland
Baltimore County
1000 Hilltop Circle
Baltimore, MD 21250
oates@cs.umbc.edu
Abstract—Jean is a computational account of several aspects
of cognitive development, specifically, a kind of Piagetian
schema-driven learning. Jean’s schemas amalgamate Piagetian
sensorimotor schemas with Image Schemas and with Dynamic
Maps. These schemas are implemented in our Image Schema
Language. Using our Experimental State Splitting algorithm,
Jean learns by composing schemas into gists and by differenti-
ating schemas. Jean works in a simple dynamical environment,
and we have started to demonstrate transfer of schemas and
gists between similar scenarios in the environment, as well as
between different environments.
Index Terms—Cognitive development, transfer learning, im-
age schemas, knowledge representation, dynamics.
I. INTRODUCTION
Jean integrates several promising ideas from cognitive
science into a model of early cognitive development. From
Piaget we borrow the ideas that children learn some of what
they know by repeatedly executing schemas, and executing
schemas is in a sense rewarding, and some new schemas are
modifications or amalgamations of old ones [1], [2]. From
the Image Schema theorists [3], [4], [5], [6], [7], [8], [9],
[10], [11] we hear that primitive schemas are encodings
or redescriptions of sensorimotor information; and these
schemas are semantically rich, general, and extend or transfer
to new situations, some of which have no salient sensorimotor
aspects. Another idea, represented by various authors, is that
semantic distinctions sometimes depend on dynamics — how
things change over time — and so schemas should have a
dynamical aspect [12], [13], [14], [15], [16], [17], [18], [19],
[20], [21], [22]. Much of our algorithmic work has been
about learning dynamical models such as maps [23], [24],
[25], finite state machines [26], other kinds of Markov chain
models [27], [28], and fluents [29], [30].
To build Jean we had to be precise about the form
and content of schemas, and about how they are learned.
One contribution of Jean is that it realizes schemas and
learning methods in computer programs, particularly, our
Image Schema Language [31] and the Experimental State
Splitting algorithm. The project is less than a year old, so
it provides more conjectures than results, but the conjectures
are worth stating: It will be possible to provide a relatively
small, core set of schemas and a general algorithm to learn
others as they are needed or indicated by experience. We are
betting on a compositional account of knowledge, in which
newly learned stuff is assembled from previously learned and
appropriately modified components. Schemas will have to
be more than the declarative, logical structures proposed by
AI researchers over the decades; they will have to include
behavior-generating controllers, dynamic maps, deictic vari-
able bindings, and causal theories; these components will
not all develop simultaneously. The transfer of schemas
between situations will prove to be the most significant
research challenge, and some transfer will be initiated by
pre-schematic, amodal representations.
II. THE IMAGE SCHEMA LANGUAGE
Image schemas are representations that are “close” to
perceptual experience. They are sometimes presented as as
re-descriptions of experience. Their popularity is due to their
supposed generality and naturalness: So many situations are
naturally described in terms of paths, up-down relations, part-
whole relationships, bounded spaces, and so on. Even non-
physical ideas, such as following an argument, containing
political fallout, and feeling “up” or “down,” seem only a
short step from image-schematic foundations [3], [4], [5].
These ideas are attractive but vague, as we discovered
when we tried to build a formal Image Schema Language
(ISL) [31]. Published accounts of schemas are often ambigu-
ous; for example, a path might be any directional sequence of
locations, or a sequence of locations that are marked by some
physical attributes, or a sequence of locations one intends to
visit, or a sequence of locations another intentional agent has
visited, and so on.
We found it necessary to distinguish three kinds of image
schema. Static schemas describe unchanging arrangements
of physical things; dynamic schemas describe how the en-
vironment changes; and action schemas describe intentional
aspects of static and dynamic schemas. Thus, the action
schema for “approaching” includes a path (a static schema)
but gives it the intentional gloss that it is the path one intends

to follow (or is following). Moving might be intentional or it
might simply be the result of force acting on an object. Both
cases involve a path, but the latter is described by a dynamic
schema, not an action schema.
All image schemas in ISL include variables that are bound
to objects in the environment. Dynamic schemas also include
maps which describe how relations (such as distance) change
over time [24], [15], [23]. Action schemas also contain
controllers that control behavior. For instance, Jean’s schema
for “approaching” binds its variables to the approaching
object (typically Jean) and the approached object, a map that
shows how distance between the objects changes over time,
and a controller that makes Jean move toward the location
of the object it is approaching. Eventually, schemas will also
be augmented with causal relations (see Sec. V).
The structure of dynamic and action schemas is described
in Section III. The rest of this section is primarily about static
image schemas.
As represented in ISL, image schemas are objects, in
the sense of the object-oriented data model. Each schema
has a set of operations that determine its capabilities. For
example, operations for a basic container schema include
putting material into a container and taking material out. Each
schema also has a set of internal slots that function as roles in
a case grammar sense [32]. Slots permit image schemas to be
related to each other through their slot values. For example,
the contents of a container can be other image schemas.
To construct more complex schemas from simpler ones,
Jean has a mechanism called interpretation, which, in object-
oriented terms is like an extended form of delegation.
Interpretations map from one or more specifications of a
“source” image schema to a “target” schema. For example,
we would probably first think to represent a room as a
location, or as a bounded space (i.e. a region), but from a fire
marshall’s perspective it would be useful to interpret a room
as a container with a capacity of some number of people.
Interpretation gives us flexibility in evaluating the proper-
ties of some domain in terms of image schemas; different
(even conflicting) interpretations of the same situation can
be maintained simultaneously. Interpretation is also critical
to metaphorical extension and bears relations to analogical
mapping [33].
To illustrate schemas in ISL, it will be helpful to walk
through an example, which we take from our work on
representing chess patterns. Consider a chess board in which
the Black queen has the White king in check. In image
schema terms, we say that there exists a path from the
queen to the king. In ISL, we generate a path schema,
which contains a set of locations, as shown in Figure 1.
Representing a path simply as a set of locations gives us
generality, but here it’s important that the queen can traverse
the path in the situation that holds currently on the board.
This is captured by an interpretation of the path as a set
of directional linkages from each location (a source) to the
next on the path (a destination). Another piece of domain
information is that no location can be occupied by more than
one piece at a time. This is represented by an interpretation
of each location as a container with a capacity of 1. When
a piece moves to a location, the container reaches capacity
and yet another image schema, empty/full, is automatically
created, indicating that the location is full.
Fig. 1. Representing blockage in ISL.
Given these image schemas, their relationships, and the
operations that they support, it becomes possible to reason
about the situation and the possible responses White can
make to counter the threat of the queen. The check exists
because the path from the queen to the king is traversable.
Traversability for a path schema is defined, in words, as
follows: a path can be traversed when every linkage between
successive locations can be traversed. Traversability for a
linkage schema, in turn, is allowed when its source can be
entered and its destination can be exited. Basic locations have
no built-in constraints on entering and exiting, but when a
location is interpretable as a container, this changes. One
cannot add more to a container that has reached capacity.
The interpretation relationships between these schemas cause
changes to propagate outward: a full container cannot be
added to; its location cannot be entered; a directional linkage
cannot be traversed (via its source); a path cannot be traversed
(due to a non-traversable linkage). The result is a new
image schema, blockage, which is created when a container
representing a location that acts as the source of a directional
linkage in a path becomes full. The contents of the container
constitute the blocker. This structured combination of image
schemas—locations, path, linkages, blockage, and so forth—
can be stored away in memory for later retrieval, limiting the

need for a complete reconstruction of the combination from
scratch.
The ISL representation provides a description of the sit-
uation in the form of a structured combination of image
schemas. Compare this combination with how we might
describe a tactic in chess: “When an opponent’s piece puts
your king in check, you can counter by moving another
piece into its path.” The combination of schemas captures
the essence of this natural language description. The repre-
sentation is general, abstracting away the specific positions
of the pieces, the existence of other pieces, even the identity
of the attacking piece. The generality of the representation
can also be seen in that its substructure maps to other basic
concepts in chess. By using object schemas that include
information about the color of a piece, we can use the
path/linkage substructure to represent a threat of one piece
on another, when the colors of the pieces are different; if
they are the same, we can represent a defense relationship.
The representation also supports the ability to reason about
emergent structure. White might have a dozen possible moves
in the situation given in the example, but few of them will
be appropriate (or even legal). One of White’s most plausible
responses, in terms of image schemas, is to recognize that
the situation is a partial match to a blockage schema (which
does not yet exist), and that a specific response will lead
to the creation of the blockage. Rather than reasoning about
the low-level properties of individual pieces, White reasons
using tactical abstractions. Other chess concepts similarly
lend themselves to abstraction that can be naturally captured
by image schemas: application of force on the opponent’s
king (even if the king is never put in check), balance in the
distribution of pieces on the board, control of the center of
the board, and so forth. Lower-level descriptions of moves
(e.g., based on paths alone) are not inaccurate, but they fail
to capture the reasons behind the moves.
III. LEARNING: EXPERIMENTAL STATE SPLITTING
Jean learns new schemas in two ways, by composing
schemas and by differentiating states. Both are accomplished
by the Experimental State Splitting (ESS) algorithm. We first
describe the data structures ESS produces and operates over,
and then describe how they are learned. Given some new en-
vironment, ESS tries to construct a finite state machine model
of the environment, using schemas as the basic building
blocks for describing states and actions. ESS differentiates,
or splits, states by identifying predictive patterns of schema
instantiation observed while interacting with the environment.
The agent’s model of the world is extended by creating new
states that incorporate these schemas. The execution of action
schemas transitions the agent between states. The resulting
finite state machine can be solved for a policy to attain some
chosen goal state in the state space. We call these learned
machines for achieving specific goals gists. Gists may, in
turn, be used as composite actions in future state machines
when the agent is faced with a new environment.
Figure 2 illustrates a gist for approaching and contacting a
cat. In the scenario this gist was learned, the cat is animate,
capable of producing, at will, a number of actions including
sitting still, walking or running away. At the same time, the
cat does respond to the distance and movement of Jean. In
particular, if Jean moves toward the cat rapidly, the cat will
run away; if Jean approaches slowly, the cat will tend to keep
doing what it is doing. Because of these behaviors, there
is uncertainty in Jean’s representation of what the cat will
do, but the general rule about how to approach the cat is
represented in the gist. Namely, the only way to catch the
cat is to first get into state s2, where the cat is nearby and
not moving quickly, and then to move fast toward the cat,
reaching state s1. All other patterns of movement leave the
robot in states s3 or s4. This corresponds to the strategy of
slowly sneaking up to the cat and then quickly pouncing on
it to catch it.
State s3 of Figure 2 is comprised of two ISL schemas:
an object schema that binds its deictic variable to the cat,
and a near-far schema that binds its two deictic variables
to the robot (i.e., Jean) and the cat, respectively. The near-
far schema also asserts that the cat is more than six units
away from the robot. s3 is associated with two action
schemas, fast-approach-object (F ) and slow-approach-object
(S), represented as arcs leaving s3. Not pictured in the figure,
but part of the representation of the approach action schemas,
is the dynamic map that describes the dynamics of distance
as one entity approaches another; we will describe the map
in a moment.
Fig. 2. A learned composite action schema for catching a cat
Jean learns gists by interacting with objects in a simulated
world. To date, Jean has learned gists for interacting with

walls (which are inanimate and static), balls (inanimate yet
dynamical), and cats (animate, intentional and dynamical).
To learn a gist like the one in Figure 2, Jean repeatedly
retrieves action schemas from its memory, runs the associated
controllers, producing actions, specifically slow and fast
movement to a location; assesses the resulting states, and, if
the transitions between states are highly unpredictable, Jean
splits states to make the resulting states more predictable.
In fact, the three states, s2, s3 and s4 were all originally
one undifferentiated state in which Jean moved either fast or
slowly toward the cat. Its inability to predict whether it could
catch the cat drove Jean to differentiate states, resulting in
the gist we have been discussing.
The ESS algorithm can be described in familiar policy-
learning terms. We assume only that Jean has a goal state
and an initial non-goal state. The following outlines the ESS
procedure:
• Initialize with two states in our state space, S0 =
{s0, sg}, where sg is the given goal state.
• While -optimal policy not found:
– Gather experience: accumulate data for the transi-
tion probabilities p(si, aj , sk).
– Find state schema si or action schema ai(si) to
split that maximizes the entropy score reduction of
the split: H(si, ai) −min(H(sk1 , ai),H(sk2 , ai)),
where sk1 and sk2 result from splitting si, or
H(si, ai)−min(H(si, ak1),H(si, ak2)), where ak1
and ak2 result from splitting ai.
– Split si ∈ St into sk1 and sk2 , and replace si with
new states in St+1, or split ai into ak1 and ak2 and
replace ai with new actions in A(si).
– Continue to gather experience, and re-solve for
optimal plan in St+1
St is the entire state space at time t. A is the set of all
actions, but A(s) ⊂ A are the actions that are valid for state
s ∈ S. A(s) should be much smaller than A. H(si, aj) is
the boundary entropy of a state-action pair, where the next
observation is one of states in St. A small boundary entropy
corresponds to a situation where executing action aj from
state si is highly predictive of the next observed state. Finally,
p(si, aj , sk) is the probability that taking action aj from state
si will lead us to state sk.
Jean searches for schemas or state variables that predict
state transitions, and uses these to split states. Figure 3
shows the distinctions that Jean draws in the process of state
splitting. First, the algorithm begins with an undifferentiated
non-goal state. Then, it learns that the type of object is
an important predictor of whether or not it can catch the
object. Balls are easy to catch, whereas cats are hard to catch.
From here, the algorithm recognizes that distance is also an
important factor in determining whether or not it can catch
a cat. When it is near the cat, executing a fast-approach-
object (F ) will frequently lead to success in catching the
cat, whereas when it is fara way from the cat, F does not
usually result in catching the cat. Thus, the algorithm splits
on distance with a threshold of 6, where <= 6 is considered
near, and > 6 is far. Finally, the algorithm may notice that
even when Jean is near the cat, sometimes it does not succeed
in catching the cat. This might be because the cat is already
moving away from the agent with some speed. Thus, ESS
may do a final split based on the velocity of the cat. This
process leads to the states s2, s3, s4 and s5 that we see in
Figure 2.
ESS splits states that have high boundary entropy, given
data accumulated by interacting with its environment, always
trying to find state descriptions that produce a low-entropy
sequence of actions that end in the goal state. Because time
is continuous in Jean’s environment, Jean needs a way to
find states in time series of sensor data. Usually, states are
found by recognizing them in sensor data, given schemas that
describe states, as shown earlier. But sometimes, when Jean
enters an unexplored part of its state space, and when it needs
to find schemas in a state description that serve as a basis
for state splitting, it needs to extract novel state descriptions
from sensor data. To do this, Jean uses a simple heuristic:
States change when several state variables change more or
less simultaneously. This heuristic is illustrated in Figure 4.
The upper two graphs show time series of five state variables:
headings for the robot and the cat (in radians), distance
between the robot and the cat, and their respective velocities.
The bottom graph shows the number of state variables that
change value (by a set amount) at each tick. When more than
a set number of state variables change, Jean concludes that
the state has changed. For example, between time period 6.2
and 8, Jean is approaching the cat, and the heuristic identifies
this period as one state. Then, at time period 8, several
indicators change at once, and we recognize that we are in
NIL
OBJECT BALL OBJECT CAT
NEAR-FAR ROBOT CAT: DISTANCE <= 6 NEAR-FAR ROBOT CAT: DISTANCE > 6
MOVEMENT CAT: VELOCITY <= 5 MOVEMENT CAT: VELOCITY > 5
Fig. 3. ESS uses schemas and schema properties to extend state descriptions
and create new states. This process “splits” less predictive states into two or
more states which better predictive qualities, resulting in the tree structure
shown above that describes the space of all the states that are created. Leaf
nodes correspond to the entire state space, where each leaf node is a state
which a state description that includes all of the schemas from the leaf to
the root node of the tree.

a new state, one which corresponds to the cat moving away
from Jean. The regions between these changes become the
dynamic maps associated with dynamic and action schemas,
and the active ISL schemas in these regions are bundled
together into composite dynamic and action schemas such
as “s2: Object : Cat ; Near-Far : Robot, Cat :Distance ≤ 6 ;
Movement Cat : Velocity ≤ 5.”
IV. THE ARCHITECTURE OF JEAN
The main functional components of Jean are illustrated
in Figure 5. Over time, Jean builds up a repository of
schemas and gists, all represented in the Image Schema
Language (ISL), as described above. When Jean has a goal,
such as catching a cat, it retrieves appropriate gist and
runs its controller, which means taking the actions that
produce transitions between states. The component of Jean
that runs schemas is called the behavior generator. Exercising
schemas produces sensory data, and lots of it. The job of
the interpreter is to retrieve and instantiate (i.e., bind the
deictic variables of) schemas from the repository. Obviously,
Jean will try to interpret the sensory data in terms of the
schemas in the gist it is running; for example, if it is trying
to catch a cat, then it will prefer to interpret the sensory
data as matching the states in Figure 2. Sometimes, though,
the fit is poor, and another schema in the repository does a
better job of explaining the sensory data. And sometimes, it
is necessary to construct a novel static, dynamic or action
schema as described in the previous paragraph.
The interpreter produces an interpretation, which Jean uses
to update the conditional probabilities of state transitions
within gists. This cycle of acting, interpreting, and updating
state transition probabilities can go on for a long time, but
eventually Jean decides that a state in the gist it is running
0
10
20
30
40
0 2 4 6 8 10 12 14
V
a
l
u
e
Time
Sensor readings: Distance, Velocities
-2
0
2
0 2 4 6 8 10 12 14
V
a
l
u
e
Time
Sensor readings: Contact, Robot and Cat Headings in Radians
0
2
4
0 2 4 6 8 10 12 14
N
u
m
b
e
r
Time
Segment Indicators
Fig. 4. New states are extracted by cutting multiple time series at places
where multiple state variables change simultaneously.
has high enough boundary entropy to warrant splitting. The
functional component labeled Experimental State Splitting
does this, and the new gist is stored in the gist repository. The
other components in Figure 5 are described in the following
section.
V. FUTURE WORK
Jean is very young and has had limited experience in a
single domain, interacting with simulated cats and balls. Our
first priority is to give Jean a lot more experience with other
kinds of objects in this domain. We also will add structural
features, particularly walls, to the environment.
One of the more interesting conjectures about Jean is that
it will transfer schemas from one domain to another. In our
view, transfer means using a gist in a novel situation and,
if it doesn’t work, modifying it as appropriate. This process
is reminiscent of one Piaget called accommodation. We are
currently testing transfer in the following protocol: There are
two situations, call them A and B, and a gist is learned in A,
and another gist, for a particular goal, is required in situation
B. The control condition is to learn the gist for B without
access to the gist for A. The treatment condition is to allow
Jean access to the gist for A, which, if transfer is working,
will have the following expected effects: Performance in B
may be immediately better (because the gist for A, though
not perfect, may suffice to accomplish the goal in situation
B, if only sometimes), and the gist for B will be learned
faster (because the gist for A is almost right, and so Jean
need only learn appropriate modifications). We are testing
Jean with this protocol with four A,B pairs: A is catching a
ball, B is catching a cat; A is catching a cat, B is catching
a cat in an environment that has walls; A is catching a cat,
B is ambushing a squad in a completely different domain, a
simulation of small-unit tactical warfare. If the experiments
show boosts in learning rates due to transfer, this will go
some way to providing concrete evidence for the claim that
Behavior
generator
Sensorimotor
data
Interpreter
Interpretation
Experimental
State Splitting
Causal
hypotheses
Experiment
planner
ISL
Gists
Schemas
Fig. 5. The main functional components of Jean

image schemas underlie knowledge in many domains.
Another line of development for Jean is intimated by the
word “experimental” in Experimental State Splitting, and by
the boxes labeled causal hypotheses and experiment planner
in Figure 5. State splitting finds factors that reduce the
entropy of state transitions, or conversely, increase the pre-
dictability of these transitions. Not all predictive relations are
causal, however. While the nature of causal relations is itself
a subject for discussion, one account is quite popular and
serves our purposes in the Jean project. The counterfactual
theory of causality says X causes Y iff X precedes Y, and X
and Y covary, and X is necessary to affect Y. The necessity
condition is framed as a counterfactual: ¬X → ¬Y . The
problem with this theory is that it does not distinguish true
causes from mere conditions; for instance, a wire is necessary
for electricity to travel from a light switch to a light bulb, but
we would not call a wire the cause when we turn on the light.
A heuristic to get around this is to assign counterfactually
necessary and proximal actions to X’s in causal models. Thus
flipping a light switch, being the most proximal action to
illumination, and counterfactually necessary, is a candidate
cause.
It is easy to find actions that are proximal to effects, and
to formulate counterfactuals relating these actions to effects.
These counterfactuals serve as causal hypotheses for Jean
to try to refute. At this point in the project, we have not
decided whether Jean will actively plan experiments to test
its hypotheses, or will simply match its experiences against
active hypotheses, looking opportunistically for refuting evi-
dence. In either case, Jean will develop causal models of its
action schemas, and will learn not only what works, but why
it works.
ACKNOWLEDGMENT
We would like to thank Wei Mu for help with implement-
ing portions of the Jean playpen scenario in breve.
REFERENCES
[1] J. Piaget. The Construction of Reality in the Child. New York: Basic,
1954.
[2] J. Piaget. The role of action in the development of thinking. In W. F.
Overton and J. M. Gallagher, editors, Knowledge and Development,
volume 1, pages 17–42. New York: Plenum, 1977.
[3] G. Lakoff and M. Johnson. Metaphors We Live By. University of
Chicago Press, Chicago, IL, 1980.
[4] G. Lakoff. Women, Fire and Dangerous Things. University of Chicago
Press, Chicago, IL, 1987.
[5] M. Johnson. The Body in the Mind: The Bodily Basis of Meaning,
Imagination, and Reason. University of Chicago Press, Chicago, IL,
1987.
[6] R. W. Langacker. Foundations of Cognitive Grammar, volume 1:
Theoretical Prerequisites. Stanford University Press, 1987.
[7] R. W. Gibbs and H. L. Colston. The cognitive psychological reality
of image schemas and their transformations. Cognitive Linguistics,
6(4):347–378, 1995.
[8] J. Mandler. How to build a baby: Ii. conceptual primitives. Psycho-
logical Review, 99:597–604, 1992.
[9] J. Mandler. The Foundations of Mind: Origins of Conceptual Thought.
Oxford University Press, 2004.
[10] W. Croft and D. A. Cruse. Cognitive Linguistics. Cambridge University
Press, 2004.
[11] T. Oakley. Image schema. In D. Geeraerts and H. Cuyckens, editors,
Handbook of Cognitive Linguistics. Osford University Press, 2006.
[12] F. Heider and M. Simmel. An experimental study of apparent behavior.
Americal Journal of Psychology, 57(2):243–259, 1944.
[13] E. Thelen and L. Smith. A Dynamic Systems Approach to the
Development of Cognition and Action. The MIT Press, Cambridge,
MA, 1994.
[14] T. Regier. The Human Semantic Potential: Spatial Language and
Constrained Connectionism. The MIT Press, 1996.
[15] P. R. Cohen. Maps for verbs. In Proceedings of the Information
and Technology Systems Conference, Fifteenth IFIP World Computer
Conference, 1998.
[16] P. W. Blythe, P. M. Todd, and G. F. Miller. How motion reveals
intention: Categorizing social interactions. In G. Gigerenzer, P. M.
Todd, and the ABC Research Group, editors, Simple Heuristics That
Make Us Smart, pages 257–285. Oxford University Press, New York,
1999.
[17] S. Intille and A. Bobick. A framework for recognizing multi-agent
action from visual envidence. In Proceedings of the Sixteenth National
Conference on Artificial Intelligence, pages 518–525, 1999.
[18] A. Bobick and J. Davis. The recognition of human movement using
temporal templates. IEEE Transactions on Pattern Analysis & Machine
Intelligence, 23(3), 2001.
[19] J. Siskind. Grounding lexical semantics of verbs in visual perception
using force dynamics and even logic. Journal of AI Research, 15:31–
90, 2001.
[20] L. Talmy. Toward a Cognitive Semantics, volume 1: Conceptual
Structuring Systems (Language, Speech and Communication). The
MIT Press, Cambridge, MA, 2003.
[21] C. T. Morrison, E. Cannon, and P. R. Cohen. When push comes to
shove: A study of the relation between interaction dynamics and verb
use. In Working Notes of the AAAI Spring Symposium Workshop:
Language Learning, an Interdisciplinary Perspective, 2004.
[22] P. R. Cohen, C. T. Morrison, and E. Cannon. Maps for verbs: The
relation between interaction dynamics and verb use. In Proceedings
of the Nineteenth International Conference on Artificial Intelligence
(IJCAI 2005), 1995.
[23] P. R. Cohen and T. Oates. A dynamical basis for the semantic content
of verbs. In Proceedings of the Grounding of Word Meaning: Data &
Models Workshop (AAAI-98), pages 5–8, 1998.
[24] M. T. Rosenstein, P. R. Cohen, M. D. Schmill, and M. S. Atkin. Action
representation, prediction and concepts. In AAAI Workshop on Robots,
Softbots, Immobots: Theories of Action, Planning and Control, 1997.
[25] M. T. Rosenstein. Concepts from time series. In Proceedings of the
Fifteenth National Conference on Artificial Intelligence, pages 739–
745, 1998.
[26] B. Krueger, T. Oates, T. Armstrong, P. R. Cohen, and C. Beal. Transfer
in learning by doing. In Proceedings of the International Joint
Conference on Artificial Intelligence (IJCAI-05), 2005.
[27] M. Ramoni, P. Sebastiani, and P. R. Cohen. Bayesian clustering by
dynamics. Machine Learning, 47(1):91–121, 2002.
[28] L. Firoiu and P. R. Cohen. Segmenting time series with a hybrid neural
network - hidden markov model. In Proceedings of the Eighteenth
National Conference on Artificial Intelligence, 2002.
[29] P. R. Cohen. Fluent learning: Elucidating the structure of episodes.
In Proceedings of the Fourth Symposium on Intelligent Data Analysis,
volume 2189, pages 268–277, 2001.
[30] P. R. Cohen, C. Sutton, and B. Burns. Learning effects of robot
activities using temporal associations. In The 2nd International
Conference on Development and Learning (ICDL 02), 2002.
[31] R. St. Amant., C. T. Morrison, Y. Chang, P. R. Cohen, and C. Beal.
An image schema language. Submitted to The 7th International
Conference on Cognitive Modelling (ICCM 2006), 2006.
[32] C. Fillmore. The case for case. In E. Bach and R. T. Harms, editors,
Universals in Linguistic Theory. Holt, Rinehart & Winston, London,
1968.
[33] D. Gentner and A. M. Markman. Structure mapping in analogy and
similarity. American Psychologist, 52(45–56), 1997.