Neural Network Training Using Lookup Tables and Parallel Synapse Weight Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Classical neural networks require significant time and resources for training due to their adaptive nature and the need for extensive statistical or algorithmic selection of synaptic weights to match input and output images effectively.
Innovation Solution
A neural network architecture that includes a plurality of inputs, synapses with corrective weights, and neurons, where a controller determines deviations from desired outputs and modifies these weights to minimize errors, allowing for concurrent or parallel training and recognition processes, facilitated by matrix operations and data processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If classical neural networks use extensive statistical or algorithmic selection of synaptic weights, then training accuracy is improved, but training time and resource consumption increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing correction values in lookup tables during an offline phase. These correction values are computed in advance based on desired weight adjustments, allowing the training process to directly retrieve and apply them without performing extensive real-time calculations. This resolves the contradiction by performing the computationally intensive work beforehand, enabling fast training with maintained accuracy.
Solution Approach 2:
The patent applies preliminary action by pre-calculating and storing correction values in lookup tables during an offline phase. These correction values are computed in advance based on desired weight adjustments, allowing the training process to directly retrieve and apply them without performing extensive real-time calculations. This resolves the contradiction by performing the computationally intensive work beforehand, enabling fast training with maintained accuracy.
2Measurement precision
If classical neural networks perform sequential training processes, then weight adjustment accuracy is improved, but processing speed decreases
Solution Approach 1:
The patent applies segmentation by dividing the training process into distinct functional components: weight calculation units that compute weight adjustments, lookup table units that store pre-computed correction values, and application units that implement the adjustments. This modular segmentation allows different segments to operate in parallel, maintaining accuracy while increasing processing speed through concurrent execution of multiple training operations.
Solution Approach 2:
The patent introduces lookup tables as intermediary structures that mediate between weight calculation and application. These tables store pre-computed correction values and act as an intermediate storage layer, allowing multiple processing units to simultaneously retrieve and apply corrections without interfering with each other. This intermediary mechanism enables parallel processing while preserving the precision of weight adjustments.
3Measurement precision
If neural networks use more synapses and neurons for better pattern recognition, then recognition accuracy is improved, but device complexity and resource requirements increase
Solution Approach 1:
The patent applies copying by creating multiple identical instances of simplified neural network units, each with its own lookup table and weight calculation capabilities. Instead of using a single complex network with many interconnected synapses, the system uses numerous independent, simpler units that can be trained and applied in parallel. This copying approach maintains or improves recognition accuracy through ensemble effects while reducing the complexity of individual network components.
Data Source
AI summary
A neural network includes inputs for receiving input signals, and synapses connected to the inputs and having corrective weights organized in an array. Training images are either received by the inputs as an array or codified as such during training of the network. The network also includes neurons, each having an output connected with at least one input via one synapse and generating a neuron sum array by summing corrective weights selected from each synapse connected to the respective neuron. Furthermore, the network includes a controller that receives desired images in an array, determines a deviation of the neuron sum array from the desired output value array, and generates a deviation array. The controller modifies the corrective weight array using the deviation array. Adding up the modified corrective weights to determine the neuron sum array reduces the subject deviation and generates a trained corrective weight array for concurrent network training.