Neural Network Class Index Tagging via Supervised STDP
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Unsupervised learning algorithms fail to provide consistent class indices, leading to random assignment of class indices, which is undesirable in applications where classification outputs are used as inputs to downstream modules, making it difficult to build a reliable interface between unsupervised learning modules and downstream modules.
Innovation Solution
A method using supervised learning and spike-timing dependent plasticity (STDP) to tag classes by training a neural network with all-to-all plastic connections between arbitrarily indexed class neurons and output neurons, ensuring that only connections between neurons representing the same class persist, thereby achieving consistent indexing regardless of the original indexing of class neurons.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If unsupervised learning algorithms are used to separate data into classes, then data classification is achieved, but class indices are randomly assigned and inconsistent
Solution Approach 1:
The patent introduces a second neural network as an intermediary between the unsupervised learning algorithm and downstream modules. This intermediary network receives outputs from the first network and learns to map arbitrary class indices to consistent semantic labels through supervised learning, thereby resolving the inconsistency problem without modifying the original unsupervised algorithm
Solution Approach 2:
The patent transforms the class representation by changing the parameter mapping relationship. Instead of relying on the original arbitrary index assignment, the system learns a new parameter mapping that associates each class index with a consistent semantic label through supervised training, effectively reparameterizing the class representation
2Reliability
If class indices are randomly assigned by unsupervised learning, then classification separation is achieved, but reliable interface with downstream modules cannot be built
Solution Approach 1:
The second neural network serves as a mediator that translates between the arbitrary indexing system of the unsupervised module and the semantic labeling system required by downstream modules. This intermediary layer enables reliable integration without requiring changes to either the unsupervised algorithm or the downstream modules
Solution Approach 2:
The system creates a copied representation of class information in the second network that is semantically meaningful and consistent. This copied representation preserves the classification information while adding the property of consistent indexing that downstream modules require
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach ensures that class indices are consistently represented, allowing for a reliable interface between unsupervised learning modules and downstream modules, and enables static or dynamic tagging of classes, improving the reliability of classification outputs.
Implementation Method 1
A method using supervised learning and spike-timing dependent plasticity (STDP) to tag classes by training a neural network with all-to-all plastic connections between arbitrarily indexed class neurons and output neurons
Data Source
AI summary
Certain aspects of the present disclosure provide methods and apparatus for creating tags (static or dynamic) for input/output classes of a neural network model using supervised learning. The method includes augmenting a neural network model with a plurality of neurons and training the augmented network using spike timing dependent plasticity (STDP) to determine one or more tags.