Neural Network Online Learning via Unified Forward-Backward Propagation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial neural networks face limitations in the efficiency of the training process and the ability to perform continuous and online training.
Innovation Solution
A computer system comprising an artificial neural network with interconnected units that allow for asynchronous calculation of output values and perturbation values, enabling simultaneous execution and training without separate steps for forward propagation and backward error propagation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate steps for forward propagation and backward error propagation are implemented, then training accuracy is improved, but training time and operational efficiency deteriorate
Solution Approach 1:
The patent combines forward propagation and backward error propagation into a single unified computational step. The key innovation is calculating perturbation values (error signals) simultaneously during the forward pass by using the current state and target state differences, eliminating the need for a separate backward pass. This merging of operations maintains training accuracy while significantly reducing training time and enabling continuous online learning.
2Reliability
If traditional batch training with separate forward and backward passes is used, then model convergence is improved, but operational continuity and online learning capability deteriorate
Solution Approach 1:
The patent enables continuous online learning by allowing the neural network to perform both forward propagation and weight updates in a continuous stream without batch processing interruptions. The unified computational approach allows each data point to be processed and learned from immediately, maintaining model convergence while providing continuous adaptability for online learning scenarios.
Solution Approach 2:
The system dynamically adjusts weights in real-time during the forward pass based on perturbation values calculated from current and target states. This dynamic weight adjustment mechanism allows the network to adapt continuously to new data streams, enhancing online learning capability while maintaining convergence through the same computational framework that ensures reliable training.
3Measurement precision
If multiple separate processing steps are implemented, then training precision is improved, but system complexity and operational overhead increase
Solution Approach 1:
The patent merges multiple separate processing steps (forward propagation, error calculation, backward propagation, weight update) into a single unified computational operation. By calculating perturbation values and updating weights simultaneously during the forward pass, the system maintains training precision while significantly reducing system complexity and eliminating the need for separate processing pipelines.
Data Source
AI summary
Provided herein is a computer system having an artificial neural network including a plurality of units to learn a certain task. The units are directionally connected and the network as a whole is organized to generate an output based on an input. The neural network is configured to calculate an internal state value for each unit of the neural network, based on output values of the other units of the neural, weights associated with the directional connections in the neural network acting between pairs of units, and a perturbation value associated with each unit of the neural network. The output value of a subset of units of the network corresponds to an input value of the neural network and the output values of a different subset of units comprise the output of the neural network.