Negative Feedback Network for Fast Target Propagation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing training techniques for learning networks, such as backpropagation, converge slowly and provide less accurate solutions due to challenges in calculating gradients for hidden layers, and methods like difference target propagation are computationally intensive and increase memory footprint.
Innovation Solution
The implementation of a system using a learning network with a negative feedback network that rapidly determines local target outputs by perturbing output signals and updating weights based on perturbations, reducing the need for autoencoders and minimizing additional hardware and memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If backpropagation is used for training learning networks, then the training process can be implemented, but the convergence speed is slow and the number of iterations required is large
Solution Approach 1:
The patent applies feedback by using the output error to directly adjust the weights in each layer through a feedback mechanism. Instead of propagating gradients backward layer by layer as in traditional backpropagation, the system uses the final output error to compute weight adjustments for all layers simultaneously through a feedback computation process, significantly accelerating convergence
Solution Approach 2:
The patent performs preliminary action by pre-computing the feedback signals and weight adjustments before actual training iterations begin. The system calculates the feedback matrices and prepares the weight update rules in advance, allowing the training process to converge faster without requiring multiple iterative passes through the entire network
2Measurement precision
If difference target propagation is used to determine local targets, then the training can proceed, but the computational complexity increases and memory footprint expands due to requiring autoencoders
Solution Approach 1:
The patent extracts the essential function of determining local targets from the complex autoencoder structure used in difference target propagation. By removing the autoencoder component and using only the necessary feedback computation based on output errors, the system achieves the same local target determination with significantly reduced computational complexity and memory requirements
Solution Approach 2:
Instead of using autoencoders to learn inverse mappings, the patent creates a simplified copy of the forward propagation computation in reverse, using the same weight matrices to compute feedback signals. This approach replicates the essential functionality without the overhead of learning separate autoencoder parameters, reducing memory footprint while maintaining accuracy
Data Source
AI summary
A system including inputs, outputs, and a learning network between the inputs and outputs is described. The learning network includes layers, each of which includes a weight layer including weights coupled with an activation layer configured to apply activation function(s). Connections are between the layers. The system also includes a negative feedback network selectively couplable between the outputs and the inputs. The weights are configured to be trained by providing to the inputs input signals corresponding to a target output, measuring output signals at the outputs with the negative feedback network decoupled, perturbing the output signals by perturbations with the negative feedback network coupled, measuring corresponding perturbations for the connections, and updating the weights based on the corresponding perturbations. The perturbations are based on a difference between the output signals and the target output.