Analogical Concept Memory for
Architectures Implementing the Common Model of Cognition

In the past two decades, algorithmic research in Soar has augmented the architecture with decalartive long-term memories (dLTMs). Soar has two - semantic (Derbinsky et al., 2010) and episodic (Derbinsky and Laird, 2009) - that serve distinct cognitive functions following the hypotheses about organization of memory in humans (Tulving and Craik, 2005). Semantic memory enables enriching what is currently observed in the world with what is known generally about it. For example, if a dog is observed in the environment, for certain types of tasks it may be useful to elaborate that it is a type of a mammal. Episodic memory gives an agent a personal history which can later be recalled to establish reference to shared experience with a collaborator, to aid in decision-making by predicting the outcome of possible courses of action, to aid in reasoning by creating an internal model of the environment, and by keeping track of progress on long-term goals. The history is also useful in deliberate reflection about past events to improve behavior through other types of learning such as reinforcement learning or explanation-based learning. Using dLTMs in Soar agents has enable reasoning complexity that wasn’t possible earlier (Xu and Laird, 2010; Mohan and Laird, 2014; Kirk and Laird, 2014; Mininger and Laird, 2018).

However, a crucial question remains unanswered - how is general world knowledge in semantic memory acquired? We posit that this knowledge is acquired in two distinctive ways. Kirk and Laird (2014, 2019) explore the view that semantic knowledge is acquired through interactive instruction when natural language describes relevant declarative knowledge. An example concept is the goal of tower-of-hanoi a small block is on a medium block and a large block is below the medium block. Here, the trainer provides the definition of the concept declaratively which is later operationalized so that it can be applied to recognize the existence of a tower and in applying actions while solving tower-of-hanoi. In this paper, we explore an alternative view that that this knowledge is acquired through examples demonstrated as a part of instruction. We augment Soar dLTMs with a new concept memory that aims at acquiring general knowledge about the world by collecting and analyzing similar experiences, functionally bridging episodic and semantic memories.

To design the concept memory, we leverage the computational processes that underlie analogical reasoning and generalization in the Companions cognitive architecture - the Structure Mapping Engine (SME; Forbus et al. 2017) and the Sequential Analogical Generalization Engine (SAGE; McLure et al. 2015). Analogical matching, retrieval, and generalization is the foundation of the Companions Cognitive architecture. In Why we are so smart?, Gentner claims that what makes human cognition superior to other animals is “First, relational concepts are critical to higher-order cognition, but relational concepts are both non-obvious in initial learning and elusive in memory retrieval. Second, analogy is the mechanism by which relational knowledge is revealed. Third, language serves both to invite learning relational concepts and to provide cognitive stability once they are learned” (Gentner, 2003). Gentner’s observations provide a compelling case for exploring analogical processing as a basis for concept learning. Our approach builds on the analogical concept learning work done in Companions (Hinrichs and Forbus, 2017). Previous analogical learning work includes spatial prepositions Lockwood (2009), spatial concepts (McLure et al., 2015), physical reasoning problems (Klenk et al., 2011), and activity recognition (Chen et al., 2019). This diversity of reasoning tasks motivates our use of analogical processing to develop an architectural concept memory. Adding to this line of research, our work shows that you can learn a variety of conceptual knowledge within a single system. Furthermore, that such a system can be applied to not only learn how to recognize the concepts but also acting on them in the environment within an interactive task learning session.

Our design exploration of an architectural concept memory is motivated by the interactive task learning problem (ITL; Gluck and Laird 2019) in embodied agents. ITL agents rely on natural interaction modalities such as embodied dialog to learn new tasks. Conceptual knowledge, language, and task performance are inextricably tied - language is a medium through which conceptual knowledge about the world is communicated and learned. Task performance is aided by the conceptual knowledge about the world. Consequently, embodied language processing (ELP) for ITL provides a set of functional requirements that an architectural concept memory must address. Embedding concept learning within the ITL and ELP contexts is a significant step forward from previous explorations in concept formation. Prior approaches have studied concept formation independently of how they will be used in a complex cognitive system, often focusing on the problems of recognizing the existence of a concept in input data and organizing concepts into a similarity-based hierarchy. We study concept formation within the context of higher-order cognitive phenomenon. We posit that concepts are learned through interactions with an interactive trainer who structures a learner’s experience. The input from the trainer help group concrete experiences together and a generalization process distills common elements to form a concept definition.

Our work is implemented in Soar and consequently, brings to bear the theoretical postulates the architecture implements. More specifically, we build upon the following theoretical commitments:

1. 1.

   Diverse representation of knowledge: In the past decade, the CMC architectures have adopted the view that architectures for general intelligence implement diverse methods for knowledge representation and reasoning. This view has been very productive in not only studying an increasing variety of problems but also in integrating advances in AI algorithmic research in the CMC framework. We contribute to this view by exploring how algorithms for analogical processing can be integrated into a CMC architecture.

2. 2.

   Deliberate access of conceptual knowledge: Following CMC architectures, we assume that declarative, conceptual knowledge is accessed through deliberation over when and how to use that knowledge. The architectures incorporates well-defined interfaces i.e. buffers in working memory that contain information as well as an operation the declarative memory must execute on the information. Upon reasoning, information may be stored, accessed, or projected (described in further detail in Section 4).

3. 3.

   Impasse-driven processing and learning: Our approach leverages impasse in Soar, a meta-cognitive signal that can variably indicate uncertainty or failure in reasoning. Our approach uses impasses (and the corresponding state stack) to identify and pursue opportunities to learn.

4. 4.

   A benevolent interactive trainer: We assume existence of an intelligent trainer that adopts a collaborative goal with the learning system that it learns correct definitions of concepts. Upon being prompted, the trainer provides correct information to the learner to base its concept learning upon.

Based on these theoretical commitments, our paper contributes an integrative account of a complex cognitive phenomenon - interactive concept learning. Specifically, this paper:

1. 1.

   defines the concept formation problem within larger cognitive phenomenon of ELP and ITL;

2. 2.

   identifies a desiderata for an architectural concept memory;

3. 3.

   implements a concept memory for Soar agents using the models of analogical processing;

4. 4.

   introduces a novel process - curriculum of guided participation - for interactive concept learning;

5. 5.

   introduces a novel framework for evaluating interactive concept formation.

Our implementation is a functional (and not an architectural) integration of analogical processing in Soar’s declarative long-term memory systems. It characterizes how an analogical concept memory can be interfaced with the current mechanisms. Through experiments and system demonstration, we show that an analogical concept memory leads to competent behavior in ITL. It supports learning of diverse types of concepts useful in ITL. Learned concept representations support recognition during ELP as well as action based on those concepts during task performance. The concepts are from few examples provided interactively.

We introduce a novel interactive process for training Aileen to recognize and use novel concepts - guided participation. Guided participation sequences and presents lessons - conjoint stimuli (world and language) - to Aileen. A lesson consists of a scenario setup in Aileen’s world and an interaction with Aileen. A scenario can be a static scene when training visual and spatial concepts or a sequence of scenes when training an action concept. An interaction has a linguistic component (content) and a non-linguistic component (signal). The signal component of instruction guides reasoning in Aileen and determines how it processes and responds to the content. Currently, Aileen can interpret and process the following types of signals:

1. 1.

   inform: Aileen performs active learning. It uses all its available knowledge to process the content through indexical comprehension (Section 3). If failures occur, Aileen creates a learning goal for itself. In this goal, it uses the current scenario to generate a concrete example of the concept described in the content. This example is sent to its concept memory. If no failure occurs, Aileen does not learn from the example. Aileen learning is deliberate; it evaluates the applicability of its current knowledge in processing the linguistic content. It learns only when the current knowledge isn’t applicable, and consequently, Aileen accumulates the minimum number of examples necessary to correctly comprehend the content.

2. 2.

   verify: Aileen analyzes the content through indexical comprehension and determines if the content refers to specific objects, spatial relationships, or actions in the accompanying scenario. If Aileen lacks knowledge to complete verification, Aileen indicates a failure to the instructor.

3. 3.

   react: This signal is defined only when the linguistic content contains a reference to an action. Aileen uses its knowledge to produce an action instantiation. Upon instantiation, Aileen determines a goal state in the environment and then plans, a sequence of actions to achieve the goal state. This sequence of actions is executed in the environment.

Incorporating these variations in how Aileen responds to the linguistic content in a lesson enables flexible interactive learning. A trainer can evaluate the current state of knowledge in Aileen by assigning it verify and react lessons. While the verify lesson tests if Aileen can recognize a concept in the world, the react lesson tests if Aileen can use a known concept to guide its own behavior in the environment. Observations of failures helps the trainer in structuring inform lessons that guide Aileen’s learning. In an inform lesson, Aileen evaluates its own learning and only adds examples when necessary. Such learning strategy distributes the onus of learning between both participants. Lessons can be structured in a flexible, reactive way in real human-robot training scenarios.

We extend the concept memory desiderata originally proposed by (Langley, 1987) to enable embedding it within larger reasoning tasks, in this case ELP and ITL:

1. D0

   Is (a) architecturally integrated and (b) uses relational representations.

2. D1

   Can represent and learn a diverse types of concepts. In particular, for Aileen, the concept memory must be able to learn visual concepts, spatial concepts, and action concepts.

3. D2

   Learn from exemplars acquired through experience in the environment. Aileen is taught through lessons that have two stimuli - a scenario and linguistic content that describes it.

4. D3

   Enable incremental accumulation of knowledge. Interactive learning is a distinctive learning approach in which behavior is intertwined with learning. It has been previously argued that interleaving behavior and learning splits the onus of learning between the instructor and the learner such that the instructor can observe the learner’s behavior and provide more examples/instruction if necessary.

5. D4

   Learn from little supervision as realistically humans cannot provide a lot of examples.

6. D5

   Facilitate diverse reasoning over definitions of concepts.

   1. (a)

      Evaluate existence of a concept in the current environment, including its typicality. This enables recognizing a concept in the environment.

   2. (b)

      Envision a concept by instantiating it in the current environment. This enables action in the environment.

   3. (c)

      Evaluate the quality of concept definitions. This enables active learning - if the quality of a concept is poor, more examples can be added to improve it.

When Aileen identifies a new concept in linguistic content (word red), it creates a new symbol RRed. This new symbol in incorporated in a map in Soar’s semantic memory and is passed on to the concept memory for creation of a new concept via the create command. The concept memory creates a new reasoning symbol as well as a new generalization context (shown in Figure 3). A generalization context is an accumulation of concrete experiences with a concept. Each generalization context is a set of individual examples and generalizations.

After creating a new concept, Soar stores an example in the concept memory. The command {store: [(isa o2 CVRed) (isa o2 CVCylinder) (isa o2 RRed)], concept: RRed} stores that the object o2 in the world is an example of the concept RRed. This example A is stored in the RRed generalization context as is - as a set of facts. Assume that at a later time, Soar sends another example B of RRed concept through the command {store: [(isa o3 CVRed) (isa o3 CVCube) (isa o3 RRed)], concept: RRed}. The concept memory adds the new example to the RRed generalization context by these two computational steps:

1. 1.

   SME performs an analogical match between the two examples. The result of analogical matching has two components: a correspondence set and a similarity score. A correspondence set contains alignment of each fact in one example with at most one fact from other. The similarity score indicates the degree of overlap between the two representations. In the two examples A and B, there are two corresponding facts: (isa o2 CVRed) aligns with (isa o3 CVRed) and (isa o2 RRed) aligns with (isa o3 RRed). If the similarity score exceeds an assimilation threshold (Table 2), SAGE continues to the next step to create a generalization.

2. 2.

   SAGE assimilates the two examples A and B into a generalization (e.g. Figure 3). It :

   1. (a)

      Uses the correspondence to create abstract entities. In the two examples provided, (isa o2 RRed) aligns with (isa o3 RRed) and (isa o2 CVRed) with (isa o3 CVRed). Therefore, identifiers o2 and o3 can be replaced with an abstract entity (GenEntFn 0 RRedMt).

   2. (b)

      Maintains a probability that a fact belongs in the generalization. Because (isa (GenEntFN 0 RRedMT) RRed) and (isa (GenEntFn 0 RRedMT) CVRed) are common in both examples, they are assigned a probability of $1$. Other facts are not in the correspondences and appear in $1$ of the $2$ examples in the generalization resulting in a probability of $0.5$. Each time a new example is added to this generalization, the probabilities will be updated the reflect the number of examples for which the facts were aligned with each other.

Upon storage in a generalization context, a generalization becomes available for matching and possible assimilation with future examples enabling incremental (D3), example-driven (D2) learning.

During indexical comprehension, Aileen evaluates if a known concept exists in the current world through the query command. Assume that in an example scene with two objects, indexical comprehension attempts to find the one that is referred to by red through {query: {scene: [(isa o4 CVRed) (isa o4 CVBox) (isa o5 CVGreen) (isa o2 CVCylinder)], pattern: (isa ?o RRed)}}. In response to this command, the concept memory evaluates if it has enough evidence in the generalization context for RRed to infer (isa o2 RRed). The concept memory performs this inference through the following computations.

1. 1.

   SME generates a set of candidate inferences. It matches the scene with the generalization in Figure 3 (right). This match results in a correspondence between the facts (isa o4 CVRed) in scene) and (isa (GenEntFn 0 RRedMt) CVRed), which aligns o4 with (GenEntFn 0 RRedMt). Other facts that have arguments that align, but are not in the correspondences, are added to the set of candidate inferences. In our example, a candidate inference would be (isa o4 RRed).

2. 2.

   AILEEN filters the candidate inferences based on the pattern in the query command. It removes all inferences that do not fit the pattern. If the list has an element, further support is calculated.

3. 3.

   AILEEN evaluates the support for inference by comparing the similarity score of the match to the match threshold. That is, the more facts in the generalization that participate in the analogical match then it is more likely that the inference is valid.

Through queries to the concept memory and resultant analogical inferences, the working memory graph (of the world in Figure 4) is enhanced. This enhanced working memory graph supports indexical comprehension in Section 3. Note that the internal concept symbols in blue (such as RBlue) are generalization contexts in the concept memory that accumulate examples from training. Consequently, the ‘meaning’ of the world blue will evolve as more examples are accumulated.

In ITL, simply recognizing that an action has been demonstrated is insufficient, the agent must also be able to perform the action if directed (desiderata D5). One of the advantages of analogical generalization is that the same mechanism is used for recognition and projection. Consider the example scene Figure 1 in which the trainer asks Aileen to move the blue cone to the right of the red cylinder using the react signal. Assume that Aileen has previously seen some other examples of this action that are stored in concept memory as episodic traces (an example is shown in 1).

During indexical comprehension, Aileen performs queries to identify the blue cone, O1, and red cylinder, O2. Similarly, it maps the verb and the related preposition to RMove and RRightOf. To act, Aileen uses its concept memory to project the action through the command {project: {trace: [(H T0 (dc o1 o2)) (H T0 (e o1 o2)) (isa AileenStartTime T0) ...], concept: RMove}. A summary is shown in Figure 5 starting at T0. In response, the concept memory performs the following computations:

1. 1.

   SME generates a set of candidate inferences. SME to matches the current scene expressed as a trace against the generalization context of the action RMove. SME generates all the candidate inferences that symbolically describe the next states of the action concept.

2. 2.

   Aileen filters the candidate inferences to determine which apply in the immediate next state (shown in Figure 5). For example, the trace in the project command contains episode T0 as the AileenStartTime. The filter computation will select facts that are expected to be held in (t) and that the (after (t) T0) holds.

This retrieval is accepted by Aileen to be next desired state it must try to achieve in the environment.