Neural Network Auxiliary Path for Training-Free Data Dependency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks require a training phase and historical data to establish data dependency, which is time-consuming and inefficient.
Innovation Solution
The proposed solution involves creating a circuit with an auxiliary path that uses data-dependent weights to generate non-linearity, allowing data dependency to be established without historical data or training, by using switches to adjust the input weights in the main path based on the outputs from the auxiliary path.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional training phase is used to establish data dependency, then network learns from historical data, but it requires time and historical data which reduces efficiency
Solution Approach 1:
The patent introduces dynamic weight control where the weights of neurons in the main path are not fixed but are dynamically adjusted based on real-time outputs from the auxiliary path. This allows the network to adapt weights during operation rather than requiring a separate training phase, eliminating the need for historical training data and reducing training time while maintaining learning capability.
Solution Approach 2:
The patent extracts the training function from the traditional training phase and relocates it to the auxiliary path that operates during normal network execution. By taking out the learning mechanism from the training phase and integrating it into the operational auxiliary path, the system eliminates the need for separate training time and historical data while maintaining the ability to learn from current data streams.
2Adaptability or versatility
If activation functions are used to add nonlinearity, then neural network behavior is enabled, but the system becomes more complex
Solution Approach 1:
The auxiliary path serves multiple functions: it provides nonlinearity through its neurons, dynamically controls the weights of the main path neurons, and enables the network to learn from current data without separate training. By making the auxiliary path multi-functional, the patent achieves complex neural network behavior while managing system complexity through a unified structure rather than separate components.
Data Source
AI summary
An apparatus and method for creating data dependency in a neural network without the need for historical data or training is described. An “auxiliary path” that is adjacent to the main path of the neural network contains neurons that receive input data, creating non-linearity beyond that normally present in the network. The outputs of the neurons in the auxiliary path do not directly feed into the layers of neurons in the main path of the network as inputs, but instead are used to “adjust” the input weights to neurons in the main path by selecting which of existing, pre-determined weights are used for any given input. No training phase is required, and the weights in the network do not change, but instead existing paths are simply opened or closed to inputs depending upon the inputs, effectively altering the input weights to the neurons in the main path.