Prof. Michael G. Dyer      HomePage

Past Research Projects (Neural/Connectionist)

DiscWorld Project: Exploration of issues involved in multi-agent learning and distributed coordination to perform construction tasks in a 2-D world consisting of same-sized, disc-shaped objects of different colors. Resulted in 2 PhDs. In the DiscoTech system, neurally controlled agents with higher-order connections learned via reinforcement, imitation, and vicarious experiences to avoid danger, find food/water, and construct structures (such as walls) by observing and inferring their parents' construction goals from parent behavior. In the ConAg system, neurally controlled agents used neural maps to act as "blueprints" that enabled them to navigate, repair, and construct arbitrary objects in a distributed fashion.

BlobsWorld Project: Acquiring language semantics via association with visual/motor sequences. All image sequences consisted of moving, mono-colored "blobs". After learning, the system (DETE) generated verbal descriptions when presented with image/motor sequences and generated image/motor sequences when given verbal sequences. For example, when shown a red ball moving slowly downward, the system generated the verbal sequence "red ball moves slowly down". Verbal, visual, and motor sequences were encoded in terms of shifted activation over dendritic compartments in maps of artificial neurons. 1 PhD completed.

Co-OpEd Project: Use of Parallel Distributed Processing (PDP) style neural architectures to perform Natural Language Processing (NLP). Resulted in 3 PhDs. The DISCERN system learned to read script-based stories and answer questions. Concurrently it learned the meanings of the words used in the stories. DISCERN made use of Simple Recurrent Neworks (SRNs), Self-Organizing Maps (SOMs) and recirculation of representations (i.e., the activation vectors used to train the networks were themselves altered over time as a side-effect of learning.). The DYNASTY system learned to read and answer questions about goal/plan-based stories. DYNASTY made use of multiple Recursive Auto Associative Memories (RAAMs) and recirculation techniques. The CRAM system read theme-based (Aesop-fable type) stories and explored the use of Tensor Networks (up to rank 5) for encoding and manipulating conceptual representations of the goal/plan errors that such stories illustrate.

Past Research Projects (Symbolic)

Edison Project: Explored issues in automated invention of simple mechanical devices and language comprehension of natural language descriptions of mechanical devices. 1 PhD completed.

HLC Project: High-level localist connectionism for natural language comprehension and generation. Resulted in 1 PhD and 1 MS. The CHIE system was a localist connectionist network that generated English and Japanese sentences from conceptual representations using spreading activation. The ROBIN system used signature activation to propagate bindings and perform inference and plan analysis necessary for word-sense disambiguation.

OpEd Project: Representation and processes for analysis of argument/belief structures in editorial text and argumentative dialogs. Resulted in 2 PhDs. The OpEd system read short editorials in the domain of economics, and answered questions about author beliefs and and attack/support relations between beliefs. The CM system employed justification structures to support analysis of arguments in an argument dialog comprehension system.

F.L.E. Project: Foundations of legal expertise and moral/ethical reasoning. This project explored cased-based, precedent-based and analogical reasoning in the domain of law and commonsense notions of fairness. Other issues included: legal language comprehension, representation of legal concepts, organization of a knowledge base of legal cases in a conceptual memory. Resulted in 1 PhD and 1 MS. The THUNDER system read ironic stories and made ethical inferences and judgements concerning the actions of narrative characters. The STARE system organized contract cases in an episodic memory and recalled prior cases in analyzing novel cases. 1 PhD completed.

Morris Project: Memory Organization and Reasoning for Invention of Stories. Resulted in 2 PhDs. The DAYDREAMER system made use of emotional and thematic structures in modeling spontaneous remindings and continuous stream of thought. The MINSTREL system generated stories that illustrated a given theme through application of author-level goals and plans.

Mentor Project: Models of intelligent advice giving. Resulted in 1 PhD and 1 MS. The AQUA system represented both tutor and learner goal/plan knowledge and used this knowledge for comparing what the tutor believes to a model of what the tutor believes the learner believes to generate more directed advice. The SHERLOCK system tutored students in the task of "graphic mapping". This is a task where the student must generat an iconic semantic network to represent the conceptual information in a fragment of text.

LWK Project: Research in language and world knowledge acquisition. Resulted in 2 PhDs. The RINA system extended its phrasal lexicon by learning phrases in context. The OCCAM system combined similarity-based learning (SBL) with explanation-based learning (EBL) to form generalizations in the domain of international economic sanctions.