Trent E. Lange

High-level Inferencing in a Connectionist Network

Connectionist models have had problems representing and applying general knowledge rules that specifically require variables. This variable binding problem has barred them from performing the high-...

Below the Surface: Analogical Similarity and Retrieval Competition in Reminding.

Abstract While the importance of surface semantic similarities between reminding cues and memory targets has been well documented, it has been unclear whether or when human memory retrieval is influenced by structural consistency (i.e., analogical similarity), the central component of analogical mapping. The ARCS model (Analog Retrieval by Constraint Satisfaction; Thagard, Holyoak, Nelson, & Gochfeld, 1990) predicts that structural consistency, as well as competition between alternative targets in memory, will indeed influence reminding. Subjects in the experiments reported here read target passages presented in an incidental learning paradigm. After a delay, subjects read other passages and wrote down any of the previously studied texts of which they were reminded by these cues. In Experiments 1-2, related cue and target sentences contained semantically similar nouns. Cue/target consistency was manipulated by varying the case-role correspondences of these nouns. In Experiment 3, related cue and target stories contained variations of the same events, and structural consistency at the level of themes was manipulated by varying the sequencing of events. In all experiments, retrieval competition was manipulated by presenting subjects with cue sentences that were related to either a single consistent or inconsistent target (singleton condition) or both a consistent and inconsistent target (competition condition). Results indicated that both retrieval competition and structural consistency influence reminding, with the impact of the latter tending to be greater in the competition condition. The pattern of results was simulated using the ARCS model. We discuss the implications of these findings for other psychological and artificial intelligence models of memory retrieval.

Remote analogical reminding

Remote analogical reminding is hypothesized to occur when one episode is cued by another sharing similar themes but not similar object, character, or event descriptions. We report three experiments exploring this view. Subjects’ remindings in Experiment 1 showed sensitivity to remote analogical similarity even though targets were encoded only briefly in an incidental learning paradigm. Experiment 2 subjects showed reliable remindings of remote analogs with study-test delays of up to 1 week. Experiment 3 demonstrated that remote analogical reminding effects are not an artifact of subjects’ editing nonanalogical remindings. All experiments supported the hypothesis that human memory is sensitive to remote analogical similarity. We discuss the implications of these findings for memory models. Future progress requires the development of formal models that quantify factors relevant to reminding performance, such as reminding interference, transfer-appropriate processing, and domain expertise.

Lexical and pragmatic disambiguation and re-interpretation in connectionist networks

Abstract Lexical and pragmatic ambiguity is a major source of uncertainty in natural language understanding. Symbolic models can make high-level inferences necessary for understanding text, but handle ambiguity poorly, especially when later context requires a re-interpretation of the input. Structured connectionist networks, on the other hand, can use their graded levels of activation to perform lexical disambiguation, but have trouble performing the variable bindings and inferencing necessary for language understanding. We have previously described a structured connectionist model, ROBIN, which overcomes many of these problems and allows the massively-parallel application of a large class of general knowledge rules. This paper describes how ROBIN uses these abilities and the contextual evidence from its semantic networks to disambiguate words and infer the most plausible plan/goal analysis of the input, while using the same mechanism to smoothly re-interpret the input if later context makes an alternative interpretation more likely. We present several experiments illustrating these abilities and comparing them to those of other connectionist models, and discuss several directions in which we are extending the model.

Analogical Retrieval Within a Hybrid Spreading-Activation Network

The mechanisms by which people are initially reminded of analogies remain poorly understood. Psychological evidence suggests that memories retrieved tend to be semantically similar and structurally consistent with the new situation. This paper describes a hybrid symbolic/connectionist model of analogical retrieval that is influenced by the language understanding process. In this model, activation is spread through a semantic network that disambiguates and makes inferences about the input cue. A network of units is then dynamically created to represent competing mappings between the cues winning interpretation and the activated subset of potential targets. Connections created from the semantic network to these new mapping units imposes priming by semantic similarity, with links between mapping units enforcing structural consistency with the cue. The mapping network then settles by constraint satisfaction into the most coherent analogical match to the cue, after which the winning mapping units feed activation back into the semantic network, where the most highly-activated target is retrieved as the analog.

Simulating Hybrid Connectionist Architectures

The symbolic and subsymbolic paradigms each offer advantages and disadvantages in constructing models for understanding the processes of cognition. A number of research programs at UCLA utilize connectionist modeling strategies, ranging from distributed and localist spreading-activation networks to semantic networks with symbolic marker passing. As a way of combining and optimizing the advantages offered by different paradigms, we have started to explore hybrid networks, i.e. multiple processing mechanisms operating on a single network, or multiple networks operating in parallel under different paradigms. Unfortunately, existing tools do not allow the simulation of these types of hybrid connectionist architectures. To address this problem, we have developed a tool which enables us to create and operate these types of networks in a flexible and general way. We present and describe the architecture and use of DESCARTES, a simulation environment developed to accomplish this type of integration.

Primary sequential memory: An activation-based connectionist model

Abstract Storage and retrieval of ordered sequences from a single, serial presentation of each element in the sequence is typically not explained by existing connectionist models. Some models finesse the issue by presenting all the elements in a sequence simultaneously. Others rely on weight-changing algorithms that require multiple presentations of the sequence. We present a model for short-term storage and recall of ordered information that relies on gated activation mechanisms. Activation from each element presented serially recruits randomly connected responder nodes whose combined activation represents the element and its position in the sequence. The sequence is later recalled by feeding activation back to the elements from the recruited responder nodes. We discuss the relevance of the model to various results from cognitive psychology, including the facts that the length-of human sequential memory is very limited, that for novel sequences recall is better for elements at the beginning and ends of sequences than for elements in the middle, and that humans have greater difficulty recalling the second occurrence of an element in a sequence containing a repeated element.