Convolutional Neural Network Twisted Interconnects for Orientation Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) face limitations in recognizing objects at orientations other than a single, default orientation, requiring additional training for varied orientations.
Innovation Solution
Implementing twisted interconnects between stages of CNNs, where data is spatially twisted, allowing the network to recognize objects at different orientations without additional training, and incorporating toroidal layers to handle data beyond array boundaries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If CNN uses standard interconnect between stages, then network structure is simple and easy to implement, but the network can only recognize objects at a single default orientation and requires additional training for varied orientations
Solution Approach 1:
The patent applies asymmetry by introducing twisted interconnects that rotate data spatially between network stages. Instead of maintaining the standard symmetric grid connectivity, the twisted interconnect rotates the data array by a twist angle (e.g., 45 degrees), creating an asymmetric connection pattern that enables the network to recognize objects at multiple orientations without additional training.
Solution Approach 2:
The patent introduces a new dimensional aspect to the network architecture by adding twist angles as an additional parameter beyond the traditional spatial dimensions. The twisted interconnect operates in a higher-dimensional space by rotating data not only in the standard x-y plane but also introducing angular transformation, effectively adding a rotational dimension to the data flow between stages.
2Reliability
If additional training is provided for varied orientations, then orientation recognition performance improves, but training time and computational resources increase
Solution Approach 1:
The patent applies preliminary action by pre-computing and hardcoding the twisted interconnect transformation matrices during network initialization. Instead of training the network to handle multiple orientations through additional data exposure, the spatial twisting operation is prepared in advance as a fixed structural component, allowing the network to immediately recognize objects at various orientations during inference without requiring extended training.
Solution Approach 2:
The patent changes the network parameters by introducing twist angles as fixed structural parameters rather than learnable weights. By setting the twist angle (e.g., 45 degrees) as a predetermined parameter in the interconnect architecture, the network achieves orientation invariance through architectural design rather than through learning additional orientation-specific patterns during training.
3Loss of information
If data array boundaries are handled with standard padding or truncation, then implementation is simple, but information at boundaries is lost or distorted
Solution Approach 1:
The patent applies the spheroidality principle by using toroidal (wrapping) boundary conditions instead of linear boundaries. The toroidal layer wraps the data array onto itself, creating a continuous surface without edges. This allows data that would normally be lost at the boundaries to be preserved by wrapping around to the opposite side, maintaining information integrity while introducing topological complexity.
Solution Approach 2:
The patent implements nesting by embedding the toroidal boundary handling within the existing convolutional layer structure. The toroidal operation is nested inside the standard convolution process, allowing boundary wrapping to be performed as an integrated part of the layer operation rather than as a separate post-processing step, thereby reducing overall system complexity.
Data Source
AI summary
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.