Data processing: measuring – calibrating – or testing – Measurement system – Measured signal processing
Reexamination Certificate
2001-09-25
2004-02-03
Hoff, Marc S (Department: 2857)
Data processing: measuring, calibrating, or testing
Measurement system
Measured signal processing
C702S125000, C702S126000, C702S178000, C702S198000
Reexamination Certificate
active
06687657
ABSTRACT:
A portion of this disclosure of this patent document contains material which is subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure, as it appears in the Patent and Trademark Office Patent files or records, but otherwise reserves all copyright rights whatsoever.
FIELD OF THE INVENTION
The present invention relates generally to a method and apparatus that senses stimuli, creates internal representations of them, and uses its sensor data and their representations to understand aspects of the nature of the stimuli (e.g., recognize them). More specifically, the present invention is a method and apparatus that represents sensed stimuli in a manner that is invariant under systematic transformations of the device's sensor states. This device need not recalibrate its detector and/or retrain its pattern analysis module in order to account for sensor state transformations caused by extraneous processes (e.g., processes affecting the condition of the device's detectors, the channel between the stimuli and the device, and the manner of presentation of the stimuli themselves).
BACKGROUND OF THE INVENTION
Most intelligent sensory devices contain pattern recognition software for analyzing the state of the sensors that detect stimuli in the device's environment. This software is usually “trained” to classify a set of sensor states that are representative of the “unknown” sensor states to be subsequently encountered. For instance, an optical character recognition (OCR) device might be trained on letters and numbers in images of printed pages. Or, a speech recognition device may be trained to recognize the spoken words of a particular speaker. After these devices have been trained, their performance may be degraded if the correspondence between the stimuli and sensor states is altered by factors extrinsic to the stimuli of interest. For example, the OCR device may be “confused” by distortions of pixel patterns due to a derangement of the camera's optical/electronic path, or it may be unfamiliar with pixel intensity changes due to altered intensity of illumination of the printed page. Similarly, the speech recognition device may be compromised if the microphone's output signal is altered by changes in the microphone's internal response characteristics, or it may fail to recognize words if the frequency spectrum of sound is altered by changes in the transfer function of the “channel” between the speaker's lips and the microphone. These processes systematically deform the sensor states elicited by stimuli and thereby define a mapping of sensor states onto one another. If such transformations map one of the sensor states in the training set onto another one (e.g., the pixel intensity pattern of one letter is mapped onto that of another letter), the pattern recognition software will misclassify the corresponding stimuli. Likewise, the device will not recognize a stimulus in the training set if it's original sensor state has been transformed into one outside of the training set.
These problems can be addressed by periodically recalibrating the device's detector to account for sensor state transformations caused by changed conditions. For example, the device can be exposed to a stimulus consisting of a test pattern that produces a known sensor state under “normal” conditions. The observed differences between the actual sensor state and ideal sensor state for this test stimulus can be used to correct subsequently encountered sensor states. Alternatively, the device's pattern analysis (e.g. pattern recognition) module can be retrained to recognize the transformed sensor states. These procedures must be implemented after each change in observational conditions in order to account for time-dependent distortions. Because the device may not be able to detect the presence of such a change, it may be necessary to recalibrate or retrain it at short fixed intervals. However, this will decrease the device's duty cycle by frequently taking it “off-line”. Furthermore, the recalibration or retraining process may be logistically impractical in some applications (e.g., computer vision and speech recognition devices at remote locations).
A similar problem occurs when the fidelity of electronic communication is degraded due to distortion of the signal as it propagates through the transmitter, receiver, and the channel between them. Most communications systems attempt to correct for these effects by periodically transmitting calibration data (e.g., test patterns) so that the receiver can characterize the distortion and then compensate for it by “unwarping” subsequently received signals. As mentioned above, these techniques may be costly because they periodically take the system “off-line” or otherwise reduce its efficiency.
SUMMARY OF THE INVENTION
The present invention substantially overcomes the disadvantages of prior sensory devices by providing a novel self-referential method and apparatus for creating stimulus representations that are invariant under systematic transformations of sensor states. Because of the invariance of the stimulus representations, the device effectively “filters out” the effects of sensor state transformations caused by extraneous processes (e.g., processes affecting the condition of the sensory device, the channel between the stimulus and the sensory device, and the manner of presentation of the stimulus itself). This means that the device can use these invariant representations to understand the nature of the stimuli (e.g., to recognize them), without explicitly accounting for the transformative processes (e.g., without recalibrating the device's detector and without retraining its pattern recognition module).
The behavior of this device mimics some aspects of human perception, which is remarkably invariant when raw signals are distorted by a variety of changes in observational conditions. This has been strikingly illustrated by experiments in which subjects wore goggles creating severe geometric distortions of the observed scene. For example, the visual input of some subjects was warped non-linearly, inverted, and/or reflected from right to left. Although the subjects initially perceived the distortion, their perceptions of the world returned to the pre-experimental baseline after several weeks of constant exposure to familiar stimuli seen through the goggles. For example, lines reported to be straight before the experiment were initially perceived to be warped, but these lines were once again reported to be straight after several weeks of viewing familiar scenes through the distorting lenses. Similar results were observed when the goggles were removed at the end of the experiment. Namely, the world initially appeared to be distorted in a manner opposite to the distortion due to the lenses, but eventually no distortion was perceived. These experiments suggest that humans utilize recent sensory experiences to adaptively “recalibrate” their perception of subsequent sensory data. There are many other examples of how our percepts are often invariant under changed observational conditions. For example, human observers are not usually confused by a different intensity of illumination of a scene. Although the raw sensory state of the observer is altered by this change, this is usually not attributed to changed intrinsic properties of the stimulus of interest (e.g., the scene). Similarly, humans perceive the information content of ordinary speech to be remarkably invariant, even though the signal may be transformed by significant alterations of the speaker's voice, the listener's auditory apparatus, and the channel between them. Yet there is no evidence that the speaker and listener exchange calibration data in order to characterize and compensate for these distortions. Rather, these observations suggest that the speech signal is redundant in the sense that listeners extract the same content from multiple acoustic signals th
Hoff Marc S
Suarez Felix
Welsh & Katz Ltd.
LandOfFree
Self-referential method and apparatus for creating stimulus... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Self-referential method and apparatus for creating stimulus..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Self-referential method and apparatus for creating stimulus... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3306350