Curvature-Selective Convolution Filters for Few-Example Visual Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current state-of-the-art neural networks lack the ability to detect higher-order features of perceptual stimuli, such as curved paths and arcs, which are easily recognized by biological neural networks, and require extensive training data.
Innovation Solution
Configure convolutional neural networks with topographically organized layers of neurons selective to contours, curve segments, and arcs without the need for training data, using excitatory and inhibitory connections to achieve selectivity for straight, curved, and arc features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional neural networks are used for visual classification, then they can achieve state-of-the-art performance on standard tasks, but they fail at tasks that are easy for infants such as learning from one or a few examples
Solution Approach 1:
The patent applies preliminary action by pre-configuring the neural network with curvature-selective convolutional kernels before training begins. These kernels are designed to detect specific curvature ranges (e.g., shallow, medium, steep curves) based on biological visual system principles. This pre-configuration provides the network with innate ability to detect curved paths and contours, enabling it to learn from fewer examples without sacrificing performance on standard tasks
2Adaptability or versatility
If standard convolutional neural networks are configured, then they can detect oriented edges, but they cannot detect higher-order features such as curved paths having a variety of arcs and sizes
Solution Approach 1:
The patent segments the curvature detection capability into multiple distinct convolutional kernels, each tuned to detect a specific curvature range. Instead of using a single generic edge detector, the system divides curvature detection into shallow curve detectors, medium curve detectors, and steep curve detectors. This segmentation allows the network to detect various higher-order features while maintaining manageable configuration complexity through systematic organization
Solution Approach 2:
The patent applies parameter changes by varying the curvature sensitivity parameters of the convolutional kernels. Each kernel is configured with specific curvature range parameters (e.g., radius of curvature, arc length) that allow it to detect particular types of curved paths. By systematically adjusting these parameters across different kernels, the network gains ability to detect diverse higher-order features including circles, arcs, and curved contours of various sizes and curvatures
3Measurement precision
If neural networks require extensive training data, then they can achieve accurate classification, but this increases loss of time and computational resources
Solution Approach 1:
The patent applies preliminary action by pre-configuring curvature-selective kernels before the training process begins. This pre-configuration provides the network with built-in expertise for detecting curved paths and contours, which are fundamental to many visual classification tasks. As a result, the network starts training with a head start, requiring fewer training iterations and less training data to achieve accurate classification, thereby reducing training time and computational resource consumption
Data Source
AI summary
Systems and methods for configuring and training neural networks for visual processing tasks, specifically focusing on higher-order feature selectivity with techniques to preconfigure higher-order features into convolutional neural networks (CNNs). The method involves configuring an artificial neural network to be selective to contours comprising curved sections and straight or nearly straight sections. This includes creating a topographically organized layer of orientation-selective neurons that collectively detect multiple orientations in an image patch. Additionally, layers of neurons selective for curve segments and approximately straight contours are created, where the selection is based on inputs from previous layers. The method further extends to creating neurons selective for curves with specified curvature, orientation, and center.