Here is a plain-text order form to order reports by mail. Some of the following reports can be downloaded as PostScript, or compressed PostScript files.
Behrooz Kamgar-Parsi and Behzad Kamgar-Parsi, "Rejection of Unacceptable Patterns with Multilayer Neural Networks," Internal Report, 1993, (NCARAI Report: AIC-93-027). Not available on-line at this time. Please see order form.
Abstract: Most of the pattern recognition applications of multilayer neural networks have been concerned with classification and not rejection of a given pattern. For example, in character recognition all alphabetical characters must be recognized as one of the 26 characters, as there is nothing to reject. However, in many situations, there is no guarantee that all the patterns that will be presented to the network would actually belong to one of the classes on which the network has been trained. In such cases, a useful network must be capable of rejection as well as classification. In this paper we propose a scheme to develop multilayer networks with rejection capabilities.
Behrooz Kamgar-Parsi and Behzad Kamgar-Parsi, "Integration of Detection and Classification of Signals Using Neural Networks," Internal Report, 1993, (NCARAI Report: AIC-93-028). Not available on-line at this time. Please see order form.
Abstract: Many pattern recognition applications involving neural networks do not deal with detection as they do only classification. In these applications, either detection is not an issue or it is done prior to classification. Indeed, the same may be true for many of the conventional approaches where detection must be completed before recognition begins. In this paper, we propose a method for the integration of detection and classification using neural networks. The method is useful for situations where unambiguous detection of a pattern or signal is not possible, e.g. in situations where we are dealing with broken or overlapping signals or signals which cannot be isolated from the background.
Behrooz Kamgar-Parsi and Behzad Kamgar-Parsi, "A Revised Clustering Technique Using a Hopfield Network," World Congress on Neural Networks, Volume IV, 24-27, Lawrence Erlbaum Associates, July 11-15, 1993, (NCARAI Report: AIC-93-030). Not available on-line at this time. Please see order form.
Abstract: Previously, Kamgar-Parsi et al. had developed an algorithm based on the Hopfield model for unsupervised clustering, or for forming self-organizing maps. Empirical results and comparisons with other techniques, including conventional techniques, have shown that the algorithm performs and scales well. Here we analyze the dynamical stability of the network by examining the eigenmodes of the connection matrix. This reveals certain shortcomings of the original formulation and a way for correcting them, which leads to a revised formulation that further improves the effectiveness of the technique. This work also signifies the importance of analyzing the connection matrix eigenmodes in designing well-behaved, stable algorithms.
Behzad Kamgar-Parsi and Behrooz Kamgar-Parsi, "The Seminal Hopfield-Tank Formulation of the Traveling Salesman Problem is Flawed," Internal Report, December 1993, (NCARAI Report: AIC-93-040). Not available on-line at this time. Please see order form.
Abstract: In neural optimization, it is essential that valid solutions be stable fixed points of the dynamics, otherwise the network will not converge to these solutions and cannot possibly find them. We prove that the original Hopfield-Tank (HT) formulation of the Traveling Salesman Problem is flawed, in that none of the valid tours are stable fixed points in the infinite gain limit. When the neuron gain is finite valid tours become only marginally stable. This helps explain the rather poor performance of the HT formulation in finding valid solutions. We also analyze the stability of several modified HT formulations, and show that some are indeed correct and effective. Empirical evidence are in agreement with our analytical results. The implication of this work is not that the Hopfield network is an inferior alternative for solving combinatorial optimizations. On the contrary, it shows that dynamical stability analysis is a tool that can help the Hopfield network realize its full potential, by identifying flaws in a heuristic formulation.
Computational Reasoning
Intelligent M4 Systems
Machine Learning
Sensor-based Systems
Back to NCARAI Library