Enhanced neural network systems and methods
Two stages of a convolutional neural network are linked by an interconnect that effects a spatial transposition of array data. The spatial transposition can include rotation, scaling, or translation (e.g., in x- or y-directions). A parameter characterizing the transposition (e.g., a parameter identifying rotation angle) can be learned by the same training process that is also used to learn other network parameters, such as layer coefficients. Additionally, or alternatively, data input to a neural network comprises—for each pixel in a patch of imagery—plural data that each indicates a relationship between the value of the pixel, and the value of a neighboring pixel. Some such neural networks can be trained to indicate the presence of a digital watermark signal in the patch of imagery—or a parameter characterizing such a digital watermark signal. Other features and arrangements are also detailed.
1. In a neural network comprising first and second stages configured for image recognition, at least one of said stages including a convolutional layer, and a method comprising the acts:
receiving an array of output data from the first stage, said array of output data having dimensions of at least 11×11 elements; and
spatially-transposing said array of output data, and employing a resulting array of spatially-transposed data as input data in the second stage;
wherein spatially-transposing the array of output data prior to provision to the second stage, and
wherein the output of the first stage is coupled to the input of the second stage by a twisted coupling interconnect to increase the reliability of image recognition at different orientations.
2. The method of claim 1 in which said spatially-transposing comprises rotating.
3. The method of claim 1 in which said spatially-transposing comprises scaling.
4. The method of claim 1 in which said spatially-transposing comprises spatial translating.
5. The method of claim 1 in which said spatially-transposing comprises rotating, scaling, and spatially-translating.
6. The method of claim 1 that includes learning a parameter that characterizes said operation through a training process.
7. The method of claim 6 in which said training process is a gradient descent training process.
8. The method of claim 6 that further includes learning coefficients for one or more of said stages by said training process.
9. The method of claim 1 in which said spatially-transposing includes computing one value in said array of spatially-transposed data as a weighted sum of plural values in said array of output data.
10. The method of claim 9 in which said array of output data includes plural rows of data including a top row, a bottom row, and plural intermediate rows, wherein said weighted sum of plural values is computed in toroidal fashion, as a sum that includes one or more values from said top row, and one or more values from said bottom row, but no value from one of said intermediate rows.
11. The method of claim 1 in which spatially-transposing the array of output data comprises one or more operations from the list: rotating, scaling and spatially-translating.
12. The method of claim 11 in which spatially-transposing the array of output data comprises two of: rotating, scaling, and spatially-translating.
13. A neural network system for image recognition, comprising:
a first stage and a second stage, at least one of said stages including a convolutional layer; and
a multi-core processor configured to:
receive an array of output data from the first stage, said array of output data having dimensions of at least 11×11 elements;
spatially-transpose said array of output data; and
employ a resulting array of spatially-transposed data as input data in the second stage,
wherein the output of the first stage is coupled to the input of the second stage by a twisted coupling interconnect to increase the reliability of image recognition at different orientations.
14. The system of claim 13 , wherein said multi-core processor is further configured to learn a parameter that characterizes said operation through a training process.
15. The system of claim 14 , wherein said training process is a gradient descent training process.
16. The system of claim 14 , wherein said multi-core processor is further configured to learn coefficients for one or more of said stages by said training process.
17. The system of claim 13 , wherein said spatially-transposing includes computing one value in said array of spatially-transposed data as a weighted sum of plural values in said array of output data.
18. The system of claim 17 , wherein said array of output data includes plural rows of data including a top row, a bottom row, and plural intermediate rows, wherein said weighted sum of plural values is computed in toroidal fashion, as a sum that includes one or more values from said top row, and one or more values from said bottom row, but no value from one of said intermediate rows.
19. The system of claim 13 , wherein the second stage comprises a digital watermarking stage, and said spatially-transposed data is utilized in the digital watermark stage to resolve image distortion.
20. A neural network system for image recognition, comprising:
a first stage, having an input and an output, for processing first input data applied to the input to produce first output data at the output, said first output data comprising array data having dimensions of at least 11×11 elements; and
a second stage having an input and an output, for processing second input data applied to the second stage input to produce second output data at the second stage output, at least one of said first stage and said second stage including a convolutional layer;
wherein the output of the first stage is coupled to the input of the second stage by a twisted coupling interconnect;
wherein said array data produced at the output of the first stage is spatially-transposed prior to being applied to the input of the second stage as the second input data to increase the reliability of image recognition at different orientations.
21. The system of claim 20 , further comprising a processor configured to learn a parameter that characterizes said operation through a training process.
22. The system of claim 21 , wherein said training process is a gradient descent training process.
23. The system of claim 21 , further comprising a processor configured to learn coefficients for one or more of said stage means by said training process.
24. The system of claim 20 , wherein said spatially transposed data is produced by computing one value as a weighted sum of plural values in said array of output data.
25. The system of claim 24 , wherein said array of output data includes plural rows of data including a top row, a bottom row, and plural intermediate rows, further comprising a processor that computes said weighted sum of plural values in toroidal fashion, as a sum that includes one or more values from said top row, and one or more values from said bottom row, but no value from one of said intermediate rows.
26. The system of claim 20 , wherein the second stage comprises digital watermarking, and said spatially-transposed data is utilized by said digital watermarking to resolve image distortion.