Computer Science Research Seminar

Fall 2004 Schedule

1. Wednesday, September 29th:
   
   Simona Doboli , Computer Science Department, Hofstra University
   
   Title: A Computational Model for Learning Unknown Sequences of Relevant Patterns Embedded in Distractors
   
   Abstract: Sequential information processing is ubiquitous throughout the nervous system. Speech recognition, language processing, spatial and motor tasks all rely on various forms of sequential processing. Sometimes, the patterns of a sequence are unknown a priori and may be interspersed with other, irrelevant information. The talk will give an overview of some of the existing neural computational models for sequence learning tasks, and will present in detail a model for discovering unknown sequences with relevant patterns embedded in distractors.

2. Wednesday, October 27th:
   
   Vincent Brown , Psychology Department, Hofstra University
   
   Title: An Associative Memory Model of Group Brainstorming
   
   Abstract: Groups and teams are popular and often-utilized in government, business, and industry for brainstorming, problem solving, and decision making. While groups may be effective in performing many tasks, laboratory evidence makes it clear that when it comes to idea generation, groups do not perform as well as an equal number of individuals brainstorming alone. There are a number of social-psychological reasons for this, including evaluation apprehension, free riding, comparison and matching processes, and production blocking. However, from a cognitive-psychological viewpoint there are theoretical reasons for believing that group brainstorming could be quite effective if the inhibitory social factors could be overcome. Computer simulations of a model based on the concept of associative memory suggest specific conditions under which brainstorming in a group could be more effective than brainstorming alone. One key aspect of the model is the notion of accessibility of ideas, which predicts that many creative ideas may be more easily generated in the context of the diverse ideas of other group members than in the familiar context of an individual's own set of ideas. Recent experimental results support many aspects of this model, which has been a useful tool for organizing and explaining much of the data on group brainstorming.

3. Wednesday, November 24th:
   
   Hua Tang , ECE Department, SUNY Stony Brook
   
   Title: Refinement based Synthesis of Continuous-Time Analog Filters Through Successive Domain Pruning, Plateau Search and Adaptive Sampling
   
   Abstract:

This presentation describes a novel algorithm for synthesis of continuous-time analog filters. The goal is to identify as many ``very good'' design points as possible without requiring feasible points as starting solutions or any other kind of additional knowledge as input. The proposed method addresses limitations of exploration based synthesis, including slow convergence and difficulties in finding diverse constraint-satisfying designs. The algorithm conducts a three step refinement process, in which poor quality solution space regions are eliminated through different strategies. It starts with the step of parameter domain pruning to identify parameter sub-domains that are more likely to produce feasible and quality solution points. Domains are found using the proposed Simplified Affine Transforms (SAT) operators. In the second step, selected variable sub-domains are searched using plateau search, a novel exploration technique described in the presentation. The algorithm addresses the three main types of solution space regions: (1) convex, quasiconvex and delta-convex regions, (2) rifts and (3) plateau. The technique expands descendant gradient based search with a systematic way of visiting plateau. The approach considers plateau as being formed of sets of equipotential points. Equipotential points are equally spaced from a certain point. As equipotential sets might become large during exploration, the method uses uniform stochastic sampling based on orthogonal arrays. Finally, the remaining regions after step two are further refined during the step of search with adaptive sampling step length. Adaptive sampling is based on the cost function gradient variation. Experiments observed the quality of results and convergence of synthesis.

BACK