Unsupervised Synaptic Weight Adjustment for Visual Cortex Feature Detectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine vision systems for image recognition and motion detection have poor recognition accuracy, while biologically plausible systems based on visual cortex techniques lack well-developed training methods due to a poor understanding of visual cortex organization and self-training methods.
Innovation Solution
A method for unsupervised training of input synapses of primary visual cortex cells and retinal ganglion cells using a simplified neural system structure, where weights are adjusted based on the sign of corresponding RGC output and activation, enabling autonomous formation of feature detectors and efficient implementation in both software and hardware.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine vision training methods are used, then training speed is fast, but recognition accuracy deteriorates (50/50 outcome for dog vs cat)
Solution Approach 1:
The system implements unsupervised self-training where neural circuits automatically organize themselves through activity-dependent plasticity rules. The neural system serves itself by using its own activity patterns to drive synaptic weight adjustments, eliminating the need for external supervised labels while achieving accurate feature detection and recognition
Solution Approach 2:
The invention changes the training paradigm from supervised parameter optimization to unsupervised parameter self-organization. By implementing activity-dependent plasticity rules that automatically adjust synaptic weights based on neural activity patterns, the system achieves both fast training convergence and high recognition accuracy through natural parameter evolution rather than iterative optimization
2Measurement precision
If biologically plausible visual cortex systems are used, then recognition accuracy improves, but training method development deteriorates (poorly developed due to poor understanding)
Solution Approach 1:
The system implements feedback loops where neural activity patterns continuously inform synaptic weight adjustments through activity-dependent plasticity rules. This biological feedback mechanism allows the system to automatically adapt and organize features without external intervention, making biologically plausible systems trainable while maintaining high recognition accuracy
Solution Approach 2:
The invention establishes preliminary organizational principles for visual cortex circuits before training begins. By pre-defining the structural framework and plasticity rules that guide self-organization, the system enables straightforward implementation of training methods while achieving accurate feature detection, thus improving ease of manufacture
3Measurement precision
If complex neural circuit structures are used, then recognition accuracy improves, but device complexity increases
Solution Approach 1:
The system segments the visual processing function into distinct neural circuit layers (e.g., retinal ganglion cells, lateral geniculate nucleus, primary visual cortex) with specialized roles. This segmentation allows accurate feature detection at each stage while keeping individual circuit modules relatively simple and manageable
Solution Approach 2:
The invention implements universal plasticity rules that apply across different neural circuit types and stages of visual processing. This multi-functionality allows the same fundamental mechanism to drive feature detection, feature refinement, and pattern recognition across the hierarchy, reducing overall design complexity while maintaining high accuracy
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Certain aspects of the present disclosure present a technique for unsupervised training of input synapses of primary visual cortex (V1) simple cells and other neural circuits. The proposed unsupervised training method utilizes simple neuron models for both Retinal Ganglion Cell (RGC) and V1 layers. The model simply adds the weighted inputs of each cell, wherein the inputs can have positive or negative values. The resulting weighted sums of inputs represent activations that can also be positive or negative. In an aspect of the present disclosure, the weights of each V1 cell can be adjusted depending on a sign of corresponding RGC output and a sign of activation of that V1 cell in the direction of increasing the absolute value of the activation. The RGC-to-V1 weights can be positive and negative for modeling ON and OFF RGCs, respectively.