Data Reprocessing and Assumption Representation in Signal Understanding Systems

A key issue in the development of next generation intelligent systems is the ability to perceive and understand the complex environments in which they will operate Complex environments are characterized by variable signal to noise ratios unpredictable source be havior and the simultaneous occurrence of target sources whose signal signatures can over lap mask or otherwise distort each other This paper argues that traditional perceptual architectures have limited e ectiveness in such environments and presents an alternative design that is a signi cant extension of the Integrated Processing and Understanding of Signals IPUS architecture The IPUS philosophy emphasizes structured bidirectional interaction between numeric signal processing and symbolic interpretation processes The interaction occurs as a result of search for signal processing control parameter values that produce evidence satisfying the interpretation processes goals This search is constrained by formal signal processing theory and dynamically generated problem solving assump tions Within the overall goal of extending generalizing and validating the IPUS architec ture this research program will explore the utility and scalability of formally designing This paper is based upon work supported by the Rome Air Development Center of the Air Force Systems Command under contract F C The content does not necessarily re ect the position or the policy of the Government and no o cial endorsement should be inferred Under development in the Computer Science Department in the University of Massachusetts at Amherst perceptual systems with three features to make their processing strategies more adaptive to complex environments demands Selective Processing Streams the ability to selectively apply signal processing algorithms to portions of the environment s signal data Multi ple Views Synthesis the ability to integrate results from several front end processings each with its own processing parameter settings and Explicit Processing Assumptions the ability to represent and revise environmental and problem solving assumptions as rst class objects and to use such information in all system components This paper presents the current status of the IPUS testbed discusses issues to be addressed in extending its architecture to accommodate the new features and describes how the new architecture should be evaluated Introduction and Motivation A key issue recognized across the machine perception subdisciplines in the development of next generation intelligent systems is the ability to perceive and understand the complex environments in which they will operate Complex envi ronments are characterized by variable signal to noise ratios unpredictable source behavior and the simultaneous occurrence of target sources whose signal signa tures can overlap mask or otherwise distort each other These observations place increased signi cance on intelligent systems perceptual components and invite a critical examination of their current design paradigm In traditional perceptual systems the front end signal processing is xed and interpretation processes are limited to analyzing only the single view a orded by that processing Fixed front end processing refers to the situation where signal processing algorithms SPAs are employed with xed control parameter values An example would be a speech recognition system whose front end processing used a xed order linear prediction algorithm with an analysis window length xed to points This traditional design paradigm is based on two assumptions A small set of front end xed parameter SPAs can produce evidence of suf cient quality for deriving plausible interpretations under all environmental scenarios A perceptual system s processing goals remain static with respect to the en vironment being monitored In restricted application domains with steady relatively high signal to noise ratios these design assumptions cause few problems When traditional perceptual systems are used to monitor complex real world environments however these assumptions lead very quickly to degraded interpretation quality The rst assumption implies that traditional perceptual systems tend to ignore shortcomings of their front end SPA and parameter sets with respect to unexpected changes in the monitored environment For example consider a system whose front end processing consists of a Fast Fourier Transform FFT with point analysis window and imagine that it is monitoring a sound source A whose two distinguishing frequency components are separated by Hz Assume that the system is supplied with a source description database that in addition to A also contains a source B with a single frequency component in the same region as source A If the signal is being sampled at KHz basic Fourier analysis theory indicates that these components separation lies at nearly twice the limit of the front end processing s frequency resolution capability If the source s components should drift in fre quency toward each other the identifying components will appear to the system s interpretation processes as one merged component which could represent source A source B or both simultaneously see gure Source behavior changes new source occurrences and other environmental events can cause SPA results to appear unlike those expected for the monitored source because the front end SPAs are not appropriate for the new environmental scenario An SPA will be termed appropriate for a set S of signals if it produces output that meets all recognition requirements e g all components adequately detected all components resolved etc for each source combination in S Systems e g that partially support the ability to recognize that signal data might have been processed by inappropriate front end SPAs may incorporate this recognition in the uncertainty associated with their output s inter pretations but they tend to have limited formal strategies to resolve this source of uncertainty As Carver points out most interpretation systems of which perceptual systems are a subset have limited means for resolving interpretation uncertainty because they are unable to generate an explicit record of the reasons for the uncertainty The rst design assumption places emphasis on developing front end SPA sets that get the right data the rst time which tends to be possible only by limiting application domains to stable constrained environments There has therefore been little motivation to provide such systems with robust capabilities to determine whether the front end SPAs are still appropriate to the environment i e detect uncertainty to explain why they are not appropriate if they have been found to be so i e generate reasons for the uncertainty and to use the explanation to modify interpretations their certainties or the front end processing itself The second assumption leads to a conceptual disconnection between the ap plication of front end SPAs and the perceptual system s dynamic processing goals It discourages the selective goal directed application of specialized often low cost SPAs to provide only enough data to resolve speci c uncertainties engendered by changes in the processing goals due to changes in the monitored signal For example a system s primary goal might be to respond to either the sounds of an infant or a ringing telephone and to ignore other sound sources This may be done by moni Partially refers to the use of catchall interpretation hypotheses such as UNKNOWN SOURCE or catchall distortion explanations such as RANDOM SOURCE NOISE or the use of clever probabilistic weighting techniques on certainty factors without explicit identi cation of source s the data could actually represent and environmental factors that led to the distortion