Neural net architecture for rate-varying inputs

• US 5,220,640 A
• Filed: 09/20/1990
• Issued: 06/15/1993
• Est. Priority Date: 09/20/1990
• Status: Expired due to Fees

• Current Assignee: Motorola, Inc. (Motorola Solutions, Inc.)
• Original Assignee: Motorola, Inc. (Motorola Solutions, Inc.)
• Inventors: Frank, Mark
• Primary Examiner(s): Fleming, Michael R.
• Assistant Examiner(s): Downs, Robert W.
• Application Number:
  US07/585,468
• Time in Patent Office: 999 Days
• Field of Search: 381/41, 381/43, 381/46, 381/47, 381/45, 364/513.5, 395/2.41
• US Class Current: 704/202
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: G06V 10/754 involving a deformation of the sample pattern or of the reference pattern; Elastic matching
• CPC Class Codes: G10L 15/16 using artificial neural networks

Abstract

A neural net architecture provides for the recognition of an input signal which is a rate variant of a learned signal pattern, reducing the neural net training requirements. The duration of a digital sampling of the input signal is scaled by a time-scaling network, creating a multiplicity of scaled signals which are then compared to memorized signal patterns contained in a self-organizing feature map. The feature map outputs values which indicate how well the scaled input signals match various learned signal patterns. A comparator determines which one of the values is greatest, thus indicating a best match between the input signal and one of the learned signal patterns.

Claims

1. A self neural net architecture for rate-varying input signals, comprising;

means for sampling a rate-varying input signal, the input signal having an initial duration, the means for sampling having an input and an output, the input of the means for sampling receiving the rate-varying input signal, the means for sampling outputting a sampled signal pattern;

means for time-scaling the sampled signal pattern, the means for time-scaling having an input and an output, the input of the means for time-scaling being coupled to the output of the means for sampling, the means for time-scaling producing a scaled signal pattern;

a feature map for comparing the scaled signal pattern to a stored signal pattern, the feature map having an input and an output, the input of the feature map being coupled to the output of the means for time-scaling; and

means for determining a correct match between the scaled signal pattern and the stored signal pattern, the means for determining a correct match having an input and an output, the input of the means for determining a correct match being coupled to the output of the feature map, the output indicating the realization of a correct match as appropriate.

2. The neural net architecture of claim 1, wherein the means for sampling comprises an analog-to-digital converter which generates digital samples of the rate-varying input signal.

3. The neural net architecture of claim 2, wherein the means for sampling further comprises a fast Fourier transform circuit coupled to the analog-to-digital converter such that the fast Fourier transform circuit receives the digital samples, the fast Fourier transform circuit transforming the digital samples from the time domain to the frequency domain.

4. The neural net architecture of claim 1, wherein the means for time-scaling comprises:

an interpolator, which expands the duration of the sampled signal pattern;

a low-pass filter which smooths the sampled signal pattern received from the interpolator; and

a decimator which compresses the time duration of the sampled signal pattern received from the low-pass filter, producing a scaled signal pattern whose net scaling is a function of the ratio of the time expansion by the interpolator to the time compression by the decimator.

5. A neural net architecture which provides for the recognition of an input signal which is a rate variant of a learned signal pattern;

comprising;

an analog-to-digital converter which creates a digital sampling of the input signal, the digital sampling having an initial duration;

a plurality of interpolators which expand the duration of the digital sampling by at least a first factor, each one of the plurality of interpolators outputting an expanded digital sampling;

a plurality of low-pass filters, each one of the plurality of low-pass filters corresponding to one of the plurality of interpolators, each one of the plurality of low-pass filters serving to smooth the expanded digital sampling output by the corresponding one of the plurality of interpolators;

a plurality of decimators, each one of the plurality of decimators corresponding to one of the plurality of low-pass filters, each one of the plurality of decimators serving to compress the duration of the expanded digital sampling smoothed by the corresponding one of the plurality of low-pass filters by at least a second factor, resulting in a plurality of scaled digital samplings;

a feature map which contains a plurality of learned signal patterns, to which the plurality of scaled digital samplings are compared, the feature map outputting a plurality of values representing how well each of the plurality of scaled digital samplings match each of the plurality of learned patterns; and

a comparator for determining which one of the plurality of values is greatest, thus indicating a best match between the input signal and one of the plurality of learned signal patterns.

6. The neural net architecture of claim 5 wherein a fast Fourier transform circuit transforms the digital sampling generated by the analog-to-digital converter from a time domain representation of the input signal to a frequency domain representation of the input signal prior to expansion by the plurality of interpolators.

7. A method for enabling a neural net feature map to recognize an input signal which is a rate variant of a learned signal pattern;

comprising;

sampling the input signal with an analog-to-digital converter, the sampled signal having an initial duration;

scaling the duration of the sampled signal by a scaling ratio by means of a method for scaling, comprising;

expanding the time duration of the sampled signal by a first factor by means of an interpolator;

filtering the expanded signal by means of a low-pass filter;

compressing the time duration of the expanded signal by a second factor by means of a decimator, providing a scaled signal, the scaling ratio being the ratio of the first factor to the second factor;

comparing the scaled signal to learned signal patterns stored in the neural net feature map; and

selecting the learned signal pattern from the neural net feature map which most closely matches the scaled signal.