US4807174A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Apparatus and method for quantizing a set of weights for a weighted sum signal processing system
Assignee: Cogent Systems Inc
•
Filing: 1985-09-27
•
Granted: 1989-02-21
•
Expires: 2005-09-27
Abstract
Abstract A conversion apparatus for quantizing a set of weights for a weighted sum signal processing system, including a predetermined number of conversion arrays operable in succession and a transfer and monitor circuit. Each conversion array stores a plurality of values to be converted, quantizes the values to be converted to produce a set of discrete values and a set of error values and replaces the contents of the means for storing with the set of discrete values. Each conversion array also combines the set of error values with a set of interpolator weights to form a transformed error set. The transfer and monitor circuit successively combines the transformed error set most recently formed with the contents of a successive conversion array, monitors the successively produced sets of error values, and halts the successive conversions and combinations when a predetermined condition is satisfied. The invention also features a method of quantizing such a set of weights.
Inventors
- 👤 John P. Costas
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-10-26 04:21
CPC Classifications
H; H03; H03H; H03H17/00; H03H17/02; H03H17/0223; H03H17/0227; G; G06; G06F; G06F7/00; G06F7/38; G06F7/48; G06F7/544; G06F7/5443
Claims Analysis
Total Claims: 31
Claim 1
Independent Claim
151 words
A conversion apparatus for quantizing a set of weights for a weighted sum signal processing system, comprising: a predetermined number of conversion arrays operated in succession, each having: means for storing a plurality of values to be converted; means for quantizing the values to be converted to produce a set of discrete values and a set of error values and for replacing the contents of said means for storing with the set of discrete values; and transformation means for combining the set of error values with a set of preselected transformation weights to form a transformed error set; and transfer means, responsive to each said conversion array, for successively combining the transformed error set most recently formed with the contents of a successive said conversion array and for monitoring the successively produced sets of error values, said transfer means halting the successive conversions and combinations when a predetermined condition is satisfied.
Claim 2
Dependent Claim
151 words
A conversion apparatus for quantizing a set of weights for a weighted sum signal processing system, comprising: a predetermined number of conversion arrays operated in succession, each having: means for storing a plurality of values to be converted; means for quantizing the values to be converted to produce a set of discrete values and a set of error values and for replacing the contents of said means for storing with the set of discrete values; and transformation means for combining the set of error values with a set of preselected transformation weights to form a transformed error set; and transfer means, responsive to each said conversion array, for successively combining the transformed error set most recently formed with the contents of a successive said conversion array and for monitoring the successively produced sets of error values, said transfer means halting the successive conversions and combinations when a predetermined condition is satisfied.
Claim 3
Dependent Claim
103 words
The apparatus of claim 1 in which each said transformation means includes: interpolation means for combining selected values with said set of interpolator weights; first means for selecting a first group of alternate values from each error value set which are combined by said interpolation means to form a first interpolated group; second means for selecting a second group of alternate error values, consisting of the remaining error values in that error value set, which are combined by said interpolation means to form a second interpolated group; and adder means for combining the first and second interpolated groups to form a transformed set.
Claim 4
Dependent Claim
14 words
The apparatus of claim 2 in which said interpolation means includes a smoothing device.
Claim 5
Dependent Claim
25 words
The apparatus of claim 2 in which each said transformation means further includes means for providing a plurality of zero values to said interpolation means.
Claim 6
Dependent Claim
27 words
The apparatus of claim 2 in which said adder means includes means for reducing by a predetermined amount the magnitude of each value of each interpolated group.
Claim 7
Dependent Claim
48 words
The apparatus of claim 2 in which each error value set has a center position and said transformation means includes means for aligning the error value set into a transformed error set having a center position corresponding in alignment with the center position of that error value set.
Claim 8
Dependent Claim
41 words
The apparatus of claim 6 in which said transfer means places the center position of the transformed error set into the same position in said successive means for storing as the position in which the previous transformed error set was placed.
Claim 9
Dependent Claim
24 words
The apparatus of claim 1 in which each said error set represents the difference between the quantized set and the set which was quantized.
Claim 10
Dependent Claim
31 words
The apparatus of claim 8 in which the contents of a storage means that is quantized, the resulting quantized set, and the resulting error set have the same number of values.
Claim 11
Dependent Claim
67 words
The apparatus of claim 1 in which said means for quantizing selects a discrete value from L cells of numerical values according to the formulas ##EQU7## where C k is the discrete value for the k th cell, B k is the lower boundary of the kth cell, P is a quantization parameter selected to provide a desired result, and L is the number of said cells.
Claim 12
Dependent Claim
19 words
The apparatus of claim 10 in which P=20 log.sub.10 (p) where p is the radix of the numerical system.
Claim 13
Dependent Claim
16 words
The apparatus of claim 11 in which the numerical system is a binary system and p=2.
Claim 14
Dependent Claim
20 words
The apparatus of claim 1 in which the contents of each said means for storing is requantized at least once.
Claim 15
Dependent Claim
27 words
The apparatus of claim 1 in which each discrete value produced by said means for quantizing is represented by a single non-zero digit in a numerical system.
Claim 16
Independent Claim
212 words
A method of quantizing an initial set of weights for a weighted-sum signal processing system comprising: quantizing the initial set of weights to obtain a first quantized set and a first error set; placing the first quantized set into a first storage means of a predetermined number of storage means; transforming the first error set into a first transformed error set; producing and entering successive quantized sets into the remaining storage means by iteratively, until a quantized set is placed into the last of the storage means, quantizing the transformed error set resulting from the most recent transformation to obtain a resulting quantized set and a resulting error set, placing the resulting quantized set into a succeeding storage means, and transforming the resulting error set into a resulting transformed error set; incrementing the most recently transformed error set into one of the storage means now occupied by a quantized set; and repeatedly, until a predetermined condition is satisfied, quantizing the contents of each successive storage means to obtain a revised quantized set and a revised error set, replacing the contents of that storage means with the revised quantized set, transforming the revised error set into a revised transformed error set, and incrementing the revised transformed error set into the next storage means.
Claim 17
Independent Claim
212 words
A method of quantizing an initial set of weights for a weighted-sum signal processing system comprising: quantizing the initial set of weights to obtain a first quantized set and a first error set; placing the first quantized set into a first storage means of a predetermined number of storage means; transforming the first error set into a first transformed error set; producing and entering successive quantized sets into the remaining storage means by iteratively, until a quantized set is placed into the last of the storage means, quantizing the transformed error set resulting from the most recent transformation to obtain a resulting quantized set and a resulting error set, placing the resulting quantized set into a succeeding storage means, and transforming the resulting error set into a resulting transformed error set; incrementing the most recently transformed error set into one of the storage means now occupied by a quantized set; and repeatedly, until a predetermined condition is satisfied, quantizing the contents of each successive storage means to obtain a revised quantized set and a revised error set, replacing the contents of that storage means with the revised quantized set, transforming the revised error set into a revised transformed error set, and incrementing the revised transformed error set into the next storage means.
Claim 18
Dependent Claim
25 words
The method of claim 15 in which quantizing provides a set of values each represented by a discrete value of a set of discrete values.
Claim 19
Dependent Claim
21 words
The method of claim 16 in which each discrete value is represented by a single nonzero digit in a numerical system.
Claim 20
Dependent Claim
63 words
The method of claim 17 in which the discrete values are selected from L cells of numerical values according to the formulas ##EQU8## where C k is the discrete value for the k th cell, B k is the lower boundary of the kth cell, P is a quantization parameter selected to provide a desired result, and L is the number of cells.
Claim 21
Dependent Claim
19 words
The method of claim 18 in which P=20 log.sub.10 (p) where p is the radix of the numerical system.
Claim 22
Dependent Claim
16 words
The method of claim 19 in which the numerical system is a binary system and p=2.
Claim 23
Dependent Claim
23 words
The method of claim 15 in which transforming includes combining the error values of each error set with a set of interpolator weights.
Claim 24
Dependent Claim
93 words
The method of claim 21 in which: a first group of alternate error values is selected from each error set and is combined with the set of interpolator weights in a first interpolation procedure to form a first interpolated group; a second group of alternate error values, consisting of the remaining error values in that error set, is selected and combined with the set of interpolator weights in a second interpolation procedure to form a second interpolated group; and the first and second interpolated groups are combined to form a transformed error set.
Claim 25
Dependent Claim
53 words
The method of claim 22 in which each error set has a center position and is transformed into a transformed error set having a center position corresponding in alignment with the center position of that error set, the transformed error set having a greater number of values distributed about the corresponding center position.
Claim 26
Dependent Claim
42 words
The method of claim 23 in which adding the transformed error set into the next storage means places the center position of the transformed error set into the same position as the position in which the previous transformed error set was placed.
Claim 27
Dependent Claim
24 words
The method of claim 15 in which each said error set represents the difference between the quantized set and the set which was quantized.
Claim 28
Dependent Claim
31 words
The method of claim 25 in which the contents of a storage means that is quantized, the resulting quantized set, and the resulting error set have the same number of values.
Claim 29
Dependent Claim
24 words
The method of claim 15 in which repeatedly quantizing, replacing, transforming and adding includes requantizing at least once the contents of each storage means.
Claim 30
Independent Claim
172 words
A method of quantizing an initial set of weight signals for a weighted-sum signal processing system comprising: quantizing the initial set of weight signals to obtain a first quantized set of signals and a first error set of signals; placing the first quantized set into a first storage means; transforming the first error set into a first transformed error set of signals; quantizing the first transformed error set to obtain a second quantized set and a second error set; placing the second quantized set into a second storage means; transforming the second error set into a second transformed error set; incrementing the second transformed error set into the first storage means; and repeatedly, until a predetermined condition is satisfied, quantizing the contents of each storage means to obtain a resulting quantized set and a resulting error set, replacing the contents of that storage means with the resulting quantized set, transforming the resulting error set into a resulting transformed error set, and incrementing the resulting transformed error set into the next storage means.
Claim 31
Independent Claim
172 words
A method of quantizing an initial set of weight signals for a weighted-sum signal processing system comprising: quantizing the initial set of weight signals to obtain a first quantized set of signals and a first error set of signals; placing the first quantized set into a first storage means; transforming the first error set into a first transformed error set of signals; quantizing the first transformed error set to obtain a second quantized set and a second error set; placing the second quantized set into a second storage means; transforming the second error set into a second transformed error set; incrementing the second transformed error set into the first storage means; and repeatedly, until a predetermined condition is satisfied, quantizing the contents of each storage means to obtain a resulting quantized set and a resulting error set, replacing the contents of that storage means with the resulting quantized set, transforming the resulting error set into a resulting transformed error set, and incrementing the resulting transformed error set into the next storage means.
References Cited (Backward Citations)
US4053754A
Cited
US4237554A
Cited
US4349889A
Cited
US4539536A
Cited
US4627072A
Cited
US4580275A
Cited
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 1992-02-21 | 1992-08-22 | 1993-02-21 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 1996-02-21 | 1996-08-22 | 1997-02-21 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2000-02-21 | 2000-08-22 | 2001-02-21 | $7700.00 | Window Closed / Lapsed |