Automated Neural Nets

Automated Target Recognition

Nabil Farhat

A dynamical approach to cognition has resulted in the first composite network for automated target recognition capable of robust distortion invariant identification of radar targets from a single echo. The network computes with diverse attractors and uses bifurcation between periodic and point attractors as the mechanism for cognition. Cognition in the context of automatic target recognition is the ability of the system, on its own, to determine that an echo belongs to a familiar (learned) target or to a novel, previously unseen, target. Cognition has important practical advantages for autonomous systems and for facilitating machine learning.

Use of Artificial Neural Network as a Risk Assessment Tool In Preventing Child Abuse

Iraj Zandi

Abstract

Every year Child Protection Services (CPS) in the US receive millions of reports of incidences of children being abused. CPS, by law, is required to investigate these reports and if necessary to intervene to protect the children at risk. Since resources are limited, some kind of triage is taking place. Various CPS offices use different risk assessment tools. The purpose of this paper is to report on a preliminary effort to explore the feasibility of utilizing artificial neural network (ANN) technology as a risk assessment tool. Third National Incidence Study of Child Abuse and Neglect (NIS-3), a congressionally mandated, periodic effort of National Center on Child Abuse and Neglect provides a rich set of data on child abuse was used to explore the utility of using ANN as a risk assessment tool. While this data set is adequate for the feasibility study, future research must rely on more targeted data.

In the current study, several different ANN designs were constructed and experimented with under a variety of conditions. These included several designs of multi layers neural network (MLNN) using several different training algorithms and a radial basis network (RBN.) The data was divided randomly to three groups: a Training Set, a Validation Set, and a Test Set. The procedure was to use the Training Set to train the network and test its veracity using the Validation Set and the Test Set.

We found that a 31-25-1 MLNN architect, after being trained, is capable of classifying the Training Set 100% accurately. However, its performance deteriorates for the Validation Set and the Test Set – new data that was not included in the Training Set. Despite this deterioration, the results are very encouraging. The network of Experiment VII-1 was capable to classify correctly 90 percent of abused cases for the Validation Set and 89 percent for the Test Set. This means that it missed 10 percent of abused children for the Validation Set (false negative) and 11 percent of the Test Set. In addition, it also misclassified 13 percent (false alarm) of those children that the survey did not find to have been abused in the Validation Set as being abused (false positive.) For the Test Set the misclassification was the same 13 percent. A radial basis network classified 100% accurately the Training Set and 93 percent of abused cases in the Test Set. It also misclassified 16 percent of not abused cases as being abused (false positive.)

Although, definitive data for the performance of currently used risk assessment tools is not available, experts tell us the performance of neural network is a clear improvement over the current practice. Much remains to be researched.